Skip to main content

Full text of "High Speed Computing Devices"

See other formats


Digitized by the Internet Archive 

in 2012 with funding from 

Gordon Bell 



http://archive.org/details/highspeedcomputiOOera 



L^ 



/ ' 






\A( 



HIGH-SPEED COMPUTING DEVICES 



HIGH-SPEED 
COMPUTING DEVICES 



By the staff of 
ENGINEERING RESEARCH ASSOCIATES, INC. 



■jf 



Supervised by 
C. B. Tompkins 

and 
J. H. Wakelin 

Edited by 
W. W. Stifler, Jr. 



First Edition 



McGRAW-HILL BOOK COMPANY, INC. 

NEW YORK TORONTO LONDON 

1950 



HIGH-SPEED COMPUTING DEVICES 

Copyright, 1950, by the McGraw-Hill Book Company, Inc. Printed in the 
United States of America. All rights reserved. This book, or parts thereof, 
may not be reproduced in any form without permission of the publishers. 



THE MAPLE PRESS COMPANY, YORK, PA. 



FOREWORD 

This volume represents the cooperative efforts of a number of 
members of the staff of Engineering Research Associates, Inc. 
C. B. Tompkins planned the report from which this volume has 
evolved and prepared much of the initial draft. The report was 
completed by the staff under the direction of J. H; Wakelin and 
was edited for publication in its present form by W. W. Stifler, Jr. 
Other contributing authors from the staff of this company were 
W. R. Boenning, W. W. Butler, A. A. Cohen, E. C. Olofson, 
L. R. Steinhardt, and Erwin Tomash. 

The contents of this volume were first assembled in the form of 
a report to the Office of Naval Research, prepared under a provi- 
sion of contract N6-ONR-240, Task 1, which called for "an 
investigation and report on the status of development of comput- 
ing machine components." Publication in the present form came 
about as a result of a suggestion of Dr. Mina Rees, then head of 
the Mathematical Sciences Division of the Office of Naval 
Research, who felt that publication over a wider range would be 
worth while. The cooperation and assistance of the Mathe- 
matical Sciences Division of ONR in bringing about the publica- 
tion of this volume is gratefully acknowledged. 

As those readers who are familiar with this field know, many 
of the large-scale computing machines already built or under 
construction have been under Government sponsorship. The 
two principal all-electronic machines now in operation, the Har- 
vard Mark III and the ENIAC, were sponsored by the Navy 
Bureau of Ordnance and the Army Ordnance Department 
respectively. The names of many of the men who have developed 
computing components and machines are mentioned in this text. 
Many of their developments have been made possible by those 
representatives of the sponsoring agencies, both civilian and mili- 
tary, whose foresight and whose confidence in the future of 
computing have made them willing to support the necessary 
research and development work. 

H. T. Engstrom 
Vice-President 
Arlington, Va. Engineering Research Associates, Inc. 

May, 1950 



PREFACE 

This volume is primarily a discussion of the mechanical devices 
and electrical circuits which can be incorporated into computing 
machines. It is not a detailed comparison of various machines. 
However, we have included descriptions of a few computers, to 
provide examples of the integration of techniques and compon- 
ents into complete systems. Because the computers built in 
this country are so much more familiar to the authors than those 
which have been built or planned in England or in other coun- 
tries, we have used American machines as examples. 

A list of all those leaders in the computing machine field with- 
out whose assistance this book could not have been written, 
would assume the proportions of a separate bibliography. We 
are therefore omitting such a list, at the same time acknowledg- 
ing with sincere thanks and appreciation the personal assistance 
of many of the men whose names do appear in the various chapter 
references and bibliography lists. 

In the preparation of this manuscript and the correction of 
the proofs the assistance of Bettie Frankl, Nancy Hall, Ann Kelley, 
and Irene Painter has been vital. We are indebted to Francis X. 
Kennelly for drawing most of the figures in this text, and to 
Harvey L. Waterman for the circuit diagrams in some of the 
earlier chapters. 

The Authors 

Arlington, Va. 
May, 1950 



Vll 



CONTENTS 

Foreword . . v 

Preface vii 

Part I. The Basic Elements of Machine Computation 

Chapter 1. Introduction 3 

References 5 

Chapter 2. Preliminary Considerations 6 

2-1. Basic Components of Automatic Computation ... 6 

Chapter 3. Counters as Elementary Components 12 

3-1. Introduction 12 

3-2. Mechanical and Electromechanical Devices .... 12 

3-3. The Flip-flop Principle 13 

3-4. Pulse-sensitive Flip-flop Circuits 17 

3-5. Radix 2 Counters 17 

3-6. The r-triode Counter, Modulo r 20 

3-7. The 2r-triode Counter, Modulo r 23 

3-8. Other Types of Flip-flop and EPDCC 25 

3-9. Economic Considerations 28 

References 29 

Chapter 4. Switches and Gates 32 

4-1. Introduction 32 

4-2. Switch Functions in Computers 32 

4-3. Types of Switches 34 

4-4. Inverse Gates, or Buffers 44 

4-5. Assemblies of Gates, Buffers, and EPDCC's ... 45 

References 54 

Chapter 5. A Functional Approach to Machine Design . . 56 

5-1. Introduction 56 

5-2. An Example 58 

ix 



x CONTENTS 

5-3. Machine Requirements 62 

5-4. Stereotyped Commands for the General-purpose 

Machine 64 

5-5. Plan for General-purpose Machine 72 

References 73 

Chapter 6. Arithmetic Systems 74 

6-1. Introduction 74 

6-2. The Fundamental Counting Systems 74 

6-3. Examples of Counting Systems 75 

6-4. Three Fundamental Theorems 77 

6-5. Rules of Arithmetic for Numbers with a Single Radix 80 

6-6. A Note Concerning Binary-to-octal Conversion . . 83 

6-7. Economy Attained by Radix Choice 84 

6-8. Negatives and Complements 87 

6-9. Scale Factors and Radical Points 89 

6-10. Binary Division — A Special Case 91 

6-11. Alternate Arrangements — Linear Mappings. ... 94 

6-12. Note on Round-off Error . . . .' 95 

6-13. Appendix 96 

References 99 

Chapter 7. Numerical Analysis 100 

7-1. Introduction 100 

7-2. The Process of Interpolation 100 

7-3. Numerical Differentiation 109 

7-4. Numerical Integration Ill 

7-5. Numerical Solution of Algebraic and Transcendental 

Equations 117 

7-6. Numerical Solution of Ordinary Differential Equa- 
tions 126 

References 130 

Part II. Computing Systems 

Chapter 8. Desk Calculators 135 

8-1. Introduction 135 

8-2. General Description of the Machines 136 

8-3. Operation of the Machines 136 

8-4. Capacity, Speed, and Capabilities 138 

8-5. Special Features 140 



CONTENTS xi 

8-6. Summary 142 

References 143 

Chapter 9. Punched-card Computing Systems 146 

9-1. Introduction 146 

9-2. The Punched Card 148 

9-3. Input 150 

9-4. Machines for Punched-card Computations 153 

9-5. Computational Operations on IBM Machines ... 160 

9-6. Transcription Devices 162 

9-7. Acquisition of Equipment 163 

9-8. Mathematical Operations to Which IBM Equipment 

Is Applicable 164 

9-9. IBM Card-programmed Electronic Calculator . . . 165 

References 165 

Bibliography 166 

Chapter 10. Large-scale Digital Computing Systems . . . 182 

10-1. Introduction 182 

10-2. Relay Computing Systems in Operation 182 

10-3. Electronic Computing Systems in Operation . . . 193 
10-4. Electronic Computing Systems in Design or under 

Development 200 

10-5. Applications of Large-scale Digital Machines . . .213 

References 218 

Chapter 11. Analog Computing Systems 223 

11-1. Introduction 223 

11-2. Components for Analog Computation 225 

11-3. The MIT Differential Analyzers 230 

11-4. The General Electric Differential Analyzer. . . . 236 

11-5. Electronic Differential Analyzer 237 

11-6. Electrical Network Analyzers 240 

11-7. Special Analog Systems 242 

11-8. Applications of Analog Computers 245 

References 245 

Chapter 12. The Form of a Digital Computer 251 

12-1. Introduction 251 

12-2. The Field of Application 251 



Xll CONTENTS 

12-3. The Estimate of Size and Speed 252 

12-4. Conclusions 260 

References 261 

Part III. Physical Components and Methods 

Chapter 13. Arithmetic Elements 265 

13-1. Introduction 265 

13-2. Series versus Parallel Operation of Arithmetic Units 265 

13-3. Operation Sequence Control 268 

13-4. Elementary Coincidence-type Adders 269 

13-5. Elementary Ring-type Adders 293 

13-6. Larger Arithmetic Units 294 

References 301 

Chapter 14. Transfer Mediums 302 

14-1. Introduction 302 

14-2. Transfer in Time 310 

14-3. Transfer in Space 374 

References 379 

Chapter 15. Data-conversion Equipment 385 

15-1. Introduction 385 

15-2. Analog-to-digital Converters 386 

15-3. Digital-to-analog Converters 393 

15-4. Conversion of Data from Written to Automatic 

Machine Form 396 

15-5. Transcription of Data from Automatic Machine to 

Written Form 404 

15-6. Radix Converters 410 

15-7. Speed-changing Devices 414 

References 416 

Chapter 16. Special Techniques and Equipment for Possible 

Use in Computing Systems 419 

16-1. Introduction 419 

16-2. New Techniques 419 

16-3. Special Electronic Tubes and Crystals 421 

16-4. Amplitude- and Frequency-discrimination Circuits 423 

16-5. Frequency-stabilizing Systems 424 

16-6. Unconventional Circuit Fabrication 424 

References 427 



CONTENTS xni 

Chapter 17. Factors Affecting Choice of Equipment. . . . 430 

17-1. Introduction 430 

17-2. Reliability of Components 430 

17-3. Possibilities for Checking 434 

17-4. Size, Cost, and Speed 439 

References 440 

Index 441 



Parti 

THE BASIC ELEMENTS OF 
MACHINE COMPUTATION 



CHAPTER 1 
INTRODUCTION 

Many of the practical problems of science, business, and war- 
fare are essentially computational. For example, scientific and 
business conclusions may be drawn by subjecting large quantities 
of observational data to appropriate statistical treatment. Mili- 
tary plans may be based on calculations from quantitative 
hypotheses involving numbers of men, numbers of items, numbers 
of distance units, and the like. The need for extensive cal- 
culations in the development of science has existed and con- 
tinually expanded. 

The existence and importance of these and an infinite number 
of other general computational problems have fostered the 
development of machine aids to computation. This volume 
touches on numerical methods and introduces the reader to 
existing computing techniques and machines. The presentation 
is intended to show how to formulate computational problems 
in ways related to the physical structure of machines. The final 
chapters list both the physical components of which the present- 
day computing instruments are comprised and other components 
most likely to find uses in the near future. A discussion of factors 
governing the choice of components is included also. 

The term component is used throughout this survey to define 
any physical mechanism or mathematical method which is used 
as a tool in automatic computation. The term is applied either 
to an abstract concept or to an item of physical equipment. 
The scope of this survey includes not only a treatment of the 
design and operation of physical mechanisms but also the arith- 
metical and analytical procedures which form the basis of solution 
of problems reduced to numerical form. 

Computational machines are classified as either digital or 
analog computers. A digital device, as the name implies, is 
one which performs mathematical operations with numbers 
expressed in the form of digits which can assume only discrete 

3 



4 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

values. The results yielded by such a device are expressed in 
digits. The precision of the computed results from a digital 
machine depends upon the number of digits it can handle, assum- 
ing that the actual operations are all performed accurately. 
(This is the equivalent of assuming that the machine is not out 
of order and that the operator makes no mistakes. A more 
detailed discussion of the concepts of precision and accuracy is 
included in Chap. 11.) 

An analog computer is one in which numbers are converted 
for purposes of computation into physically measurable quanti- 
ties such as lengths, voltages, or angles of displacement. Com- 
puted results are obtained by the interaction of moving parts or 
electrical signals related in such a manner as to solve an equation 
or perform a given set of arithmetical or mathematical operations. 

The precision of the results which can be expected from any 
given analog device, as contrasted with a digital device, depends 
upon the precision with which the device is fabricated, the skill 
and uniformity with which it is operated (including its depend- 
ence upon outside factors such as line-voltage variations), and 
the precision with which the answer can be read if the final con- 
version to digital form is made by reading a calibrated scale. 
In short, it is subject to the systematic and human errors which 
are inherent in the use of any measuring apparatus. 

A slide rule is an example of an analog computer. In this 
device, lengths correspond to the logarithms of numbers. Loga- 
rithms are added and subtracted, and numerical answers are read 
from the scale. An example of a digital computer is the abacus. 
The Chinese who uses this device assigns discrete numerical 
values to the beads and obtains precise digital answers to his 
problems by proper manipulation of the beads. 

It is hoped that those readers who are familiar with the mathe- 
matical background on which this survey has been based will 
find the descriptions of physical elements and their interrelation- 
ships of interest. For those well grounded in electrical engineer- 
ing and particularly in electronics, it is hoped that the chapters 
dealing with mathematical methods and with computing systems 
will be sufficiently straightforward to require no unreasonable 
special study for thorough understanding. 

For those readers who are particularly interested in any 
isolated phase (mathematical components, physical components, 



INTRODUCTION 5 

or computing systems), the following classification of the several 
chapters will be found useful. The chapters in each class are 
essentially independent of those in other classes. 

A. General 

Chapter 1. Introduction 

Chapter 2. Preliminary Considerations 

B. Mathematical Components 
Chapter 6. Arithmetic Systems 
Chapter 7. Numerical Analysis 

C. Physical Components and Methods 

Chapter 3. Counters as Elementary Components 
Chapter 4. Switches and Gates 
Chapter 13. Arithmetic Elements 
Chapter 14. Transfer Mediums 
Chapter 15. Data-conversion Equipment 
Chapter 16. Special Techniques and Equipment for Pos- 
sible Use in Computing Systems 
Chapter 17. Factors Affecting Choice of Equipment 

D. Computing Systems 

Chapter 5. A Functional Approach to Machine Design 

Chapter 8. Desk Calculators 

Chapter 9. Punched-card Computing Systems 

Chapter 10. Large-scale Digital Computing Systems 

Chapter 11. Analog Computing Systems 

Chapter 12. The Form of a Digital Computer 

Additional introductory material on the subject of computa- 
tion both by analog and by digital methods will be found in the 
following bibliography, containing three volumes all of which 
have been published since most of the material in the following 
pages was prepared. Hartree 2 is concerned primarily with the 
mechanisms of computing, while Berkeley 1 and Wiener 3 devote 
their efforts respectively to more popular discussion and to the 
philosophy of the subject. 

REFERENCES 

1. Berkeley, E. C, Giant Brains (John Wiley and Sons, Inc., New York, 
1949). 

2. Hartree, D. R., Calculating Instruments and Machines (University of 
Illinois Press, Urbana, 111., 1949). 

3. Wiener, Norbert, Cybernetics (John Wiley and Sons, Inc., New York, 
1948). 



CHAPTER 2 
PRELIMINARY CONSIDERATIONS 

2-1. Basic Components of Automatic Computation 

The basic components required for a general-purpose com- 
puting machine consist of (1) the input system; (2) the arithmetic 
techniques; (3) the system of operations, which includes arith- 
metic elements, storage, and control; and (4) the output system. 
Each of these components will be treated here in order to present 
an introductory view of the fundamental requirements of auto- 
matic computation. 

2-1-1. Input Systems. This is a general term for the com- 
ponents and processes by means of which problems are introduced 
to the machine. For example, the input system associated with 
a standard desk calculator is comprised of a manually operated 
keyboard of decimal digits and commands {e.g., add, multiply). 
The input more nearly defines the use to which a machine can 
be put than does any other component; there is no universally 
applicable input medium. 

Machines can be divided roughly into two classes according 
to their expected use, and each of these classes can be divided 
further into two subdivisions according to versatility: 

Class I A — Data-reduction equipment, general-purpose. 
Class IB — Data-reduction equipment, special-purpose. 
Class II A — Mathematical equipment, general-purpose. 
Class IIB — Mathematical equipment, special-purpose. 

The input requirements of these four types differ radically. 

Data-reduction machines typically receive voluminous data 
from physical measurements. Usually manual transcription 
time (the time spent in reading and transcribing the data from 
the measuring instrument) exceeds manual reduction time (the 
time spent on the ensuing arithmetic calculation). The use of 
automatic computing equipment for large-scale reduction of data 
will be strikingly successful only if means are provided for the 

6 



PRELIMINARY CONSIDERATIONS 7 

automatic transcription of these data to a form suitable for 
automatic entry into the machine. For some applications, of 
which the most prominent are those in which the reduced data 
are used to control the process being measured, the input must 
be developed for on-line operation. In on-line operation the 
input is communicated directly and without delay to the data- 
reduction device. For other applications, off-line operation, 
involving automatic transcription of data in a form suitable for 
later introduction to the machine, may be tolerated. These 
requirements may be compared with teletype operating require- 
ments. For example, some teletype machines operate on line. 
Their operators are in instantaneous communication. Other 
teletype machines are operated off line, through the intervention 
of punched paper tape. The message is preserved by means of 
holes punched in the tape and is transmitted later by feeding the 
tape to another machine. This method permits fast regular 
transmission independent of operator variations. Line time is 
thus conserved at the expense of elapsed communications time. 

The typical mathematical machine, as contrasted with the 
data-reduction machines, receives input data generated in a 
human brain. These data are not likely to be numerous, and 
automatic entry from the human brain is not presently feasible. 

2-1-2. Arithmetic Techniques. When a project involving 
extensive computation has reached the point in its development 
where the use of large-scale computing equipment is being 
planned, it is usually true that a fairly well defined field of prob- 
lems has been formulated. These are the problems the com- 
puting equipment must solve. Between the general formulation 
of this field and the production of solutions of problems from the 
field come the specific arithmetic formulations of problems and 
the development and construction of equipment which will 
yield solutions economically. Roughly, then, the project can 
be thought of as including a mathematical aspect and an engi- 
neering aspect. A precise definition of the boundary between 
these aspects is not worth attempting; one merges into the other. 
Neither part of the problem can be solved independently of the 
other. The solution must be evolved from a utilization of 
mathematical processes (including arithmetic, logic, etc.) and 
engineering processes which together yield the desired solution 
with acceptable economy. 



8 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

The central meeting ground of these two aspects of the com- 
puting program is an area which includes both the set of ele- 
mentary operations which the device is capable of performing 
and the general procedure for communicating to the device the 
set of orders it must perform in solving a problem. The term 
arithmetic techniques is used to describe this area. The term 
denotes, therefore, those arithmetic and logical operations and 
combinations of operations which are used by the machine opera- 
tor in the application to any problem of abstract arithmetic 
processes. Thus, when an operator uses a simple adding machine 
to obtain a product of two numbers, he is performing the process 
known as multiplication via the arithmetic technique of iterative 
addition. 

The speed with which a given problem can be solved with a 
particular machine depends primarily upon the number of opera- 
tions (e.g., multiplications) which the machine must perform in 
solving the problem and upon the time spent by the operator 
in preparing a program of commands ordering the machine to 
perform the required operations. For maximum versatility 
and speed, the machine must be capable of a maximum number 
of different operations. Generally speaking, the more extensive 
the variety of possible operations, the less complex will be the 
program of operations for a given problem. On the other hand, 
the variety of operations must be restricted to some extent for 
practical and economic reasons. Ingenuity in preparing pro- 
grams of operations must be substituted to some extent for 
variety of available operations. 

Therefore, in setting performance specifications for a machine, 
the choice of an optimum set of available operations depends 
upon the field of problems which the machine must solve, the 
speed with which solutions are actually required, and the prac- 
ticability of constructing physical components which will perform 
the various mathematical processes which might be desirable. 
These factors are related to the choice of an optimum set of 
operations through the arithmetic techniques which will be 
employed. 

2-1-3. System of Operations. The portion of a computer 
which actually performs the mathematical functions in the 
solution of a problem may be termed the operations system. 
It consists of arithmetic units, which perform the operations of 



PRELIMINARY CONSIDERATIONS 9 

addition, subtraction, multiplication, and division; storage 
units, which provide means for holding information for reuse; 
and control elements, which interpret commands and initiate 
arithmetic processes. 

The arithmetic units of the machine are only partially deter- 
mined by the determination of the elementary arithmetic opera- 
tions they are to carry out. The required accuracy, speed, and 
versatility of the machine all affect the design of the arithmetic 
units. If the operation has to be synchronized with outside 
phenomena, which is the case frequently in on-line operation, 
this too must be taken into account in the design of the arith- 
metic elements. Finally, it is true that the arithmetic elements 
must be designed in such a way as to permit convenient informa- 
tion transfer between the arithmetic units and the storage units 
chosen for the machine. 

It has already been stated that storage units are required to 
hold information for later reuse. This implies a concept of time 
which may be used to distinguish between types of storage: 
quick-access, or high-speed, storage; slow-access, or low-speed, 
storage; and intermediate storage. These terms, however, 
merely denote the form of the storage and omit any reference to 
the actual purpose of storage. 

A more descriptive definition of storage facilities may be 
obtained by considering the use of storage in a typical example. 
When a piece of paper, a pencil, the physical actions of writing, 
and the mental processes of calculations are involved in the 
solution of an arithmetic problem, it is readily seen that there 
are four kinds of storage required: (1) storage of the original 
data of the problem upon the paper; (2) mental storage of carry- 
overs in processes of addition, subtraction, multiplication, and 
division; (3) storage of intermediate results upon the paper; 
and (4) storage of the final solution upon the paper. Of these 
four kinds of storage, all but the second require transfer to a 
medium external to that in which the fundamental arithmetic 
processes are performed. This suggests that storage may be 
further defined as either internal or external; i.e., either for 
immediate automatic reuse in the arithmetic units, or for reuse 
in the computation process at a time determined by the pro- 
gramming of operational commands. For the former, quick 
access time is convenient; for the latter, a lower speed is per- 



10 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

missible. In fact, a lower speed may be preferable. The factors 
influencing the choice of storage speeds, the amount of storage, 
and storage mediums are more fully discussed in Chap. 14. 

The third part of the system of operations is the device which 
controls the computation process, ordering the arithmetic units 
to perform their various functions upon various numbers in 
proper sequence. These commands or orders to the arithmetic 
units may be of two types: (1) orders included in a program of 
such orders, drawn up by a human operator; or (2) orders derived 
from previous machine operations. The complexity of the 
program of orders required for the control of the machine depends 
upon the type of application for which the machine was designed. 
Consider, for example, a machine designed for solving a single 
complex problem, requiring many basic operations to be per- 
formed in a particular sequence. This is a special-purpose 
machine. In theory, a machine may be so highly specialized 
that after the values of parameters are specified, the only external 
command required is start. The rest of the commands are 
permanently built into the machine in some manner or are 
derived automatically from operations performed earlier in the 
course of the machine's solution. At the other extreme, a 
machine may be so general that an infinite variety of sequences 
of operations is possible, either by rearrangement of some 
temporary wiring system, by revision of the program encoded 
on the input medium, or by both. 

For general-purpose machines, the present tendency is to 
encode the commands in a numerical code. The machine reads 
the coded commands inscribed in proper sequence on the input 
medium. The code is usually composed of groups of digits; 
each group is composed of a part which has to do with the 
operation involved and a part which has to do with the location 
in storage of the number or numbers to be operated upon. 
There may also be a part which has to do with the choice of the 
next command to be carried out. 

2-1-4. Output System. The output problem is similar to the 
input problem. The output, as well as the input, must be care- 
fully matched to the utilization expected of the machine. If 
the instrument is to control a process, and if this process con- 
tinues during the calculation, then the output may be some sort 
of on-line device possibly including facilities for translating a 



PRELIMINARY CONSIDERATIONS 11 

digital output to continuously variable control signals. If the 
output is a printing device, it is important to avoid letting this 
printer retard the operation of the whole machine, as it will if 
the output plans are overextensive. Another consequence of 
overextensive output plans is the generation of more printed 
material than anyone can read. 



CHAPTER 3 
COUNTERS AS ELEMENTARY COMPONENTS 

3-1. Introduction 

In effect, every high-speed computing machine is composed of 
an arrangement of fundamental components. It seems reason- 
able here to study these elementary components in some detail; 
later it will be seen how they may be assembled into a complete 
equipment. It is fortunate that the number of different types 
of such fundamental components is limited. This makes it 
possible to deal with much of the subject by treating only one 
general type of elementary potential digital computing component 
(hereafter referred to as EPDCC). 

We define an EPDCC as any system, mechanical, electrical, 
electronic, electromechanical, electromagnetic, or other, which 
has the following properties: 

1. It may assume any one of a discrete set of stable states, 
the number of stable states in the discrete set being fixed 
in advance. 

2. It may influence other components, including other 
EPDCC's, or be influenced by them in different ways, 
depending upon the state in which the components are set. 

The application of these components to computing has been 
indicated in the introductory material; it will be more explicitly 
indicated in later chapters. In this chapter, discussion will be 
limited to a few elementary mechanical devices, to electrical 
circuits involving vacuum tubes, and in particular, to those 
circuits based on the flip-flop principle. 

3-2. Mechanical and Electromechanical Devices 

A common EPDCC is the 10-position adding-machine wheel 
which is used in mechanical desk computers, and which satisfies 
the definition of an EPDCC in all respects. This mechanical 

12 



COUNTERS AS ELEMENTARY COMPONENTS 13 

counter can be transformed into an electromechanical counter 
by the addition of 10 position contacts and pickup wipers. 
With this modification, the stable state of the wheel is deter- 
mined electromechanically, and the influence of the wheel on 
other components is completely electrical. An example of such 
a device is the counter developed for card-tabulating purposes. 
This device comprises a 10-position wheel, the exact position of 
which is controlled by an electromechanically operated clutch. 
The wheel position is read by means of a set of 10 contacts. 
This type of electromechanical unit also fits the definition of an 
EPDCC. Among the few existing large-scale computing 
machines there is at least one, the IBM Automatic Sequence 
Controlled Calculator presented to Harvard University in 1944 
and generally referred to as the Mark I, the entire action of which 
is based on the use of such electromechanical devices. 1 ' 13,14 

There is a fundamental limitation, however, to the maximum 
speed at which electromechanical devices can operate, owing to 
the inertia of the mechanically moving parts. Present practice 
indicates that the maximum speed for reliable operation of any 
electromechanical EPDCC is approximately 100 counts per 
second. 21 - 36 Since the great majority of future large-scale 
computing machines will demand basic speeds considerably in 
excess of this figure, it is necessary to go to some other type of 
EPDCC to achieve the desired results. The use of electron 
tubes offers such a possibility, and the remainder of this chapter 
will deal only with those EPDCC's which utilize electron tubes. 

3-3. The Flip-flop Principle 

Historically, the flip-flop principle has been known since 
1919, n but its, application to a full-scale digital computing 
machine does not appear to have been realized until 20 years 
later. 9 Prior to this conception certain special forms of flip-flop 
had been developed, but mainly for the purposes of counting 
atomic particles. Since that time the basic principle of the 
flip-flop and the ring counters has been altered variously for 
particular purposes. 

The flip-flop configuration of vacuum tubes is characterized 
by the fact that there are two plate-to-grid couplings and a 
common bias arrangement between the two halves of the circuit; 
this coupling and bias arrangement causes the circuit as a whole 



14 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

to remain in either one of its two stable states until the appli- 
cation of the signal which changes it to the other. 

Because the flip-flop circuit is so fundamental to practically 
every type of EPDCC, we shall proceed first to a description of 
this arrangement of two triodes (Fig. 3-1) having two stable 
states. 11 ' 31 ' 32 ' 34,38 - 43 In this circuit the plate of the first triode 
Vi is connected to the grid of the second V2, and the plate of the 




PULSE 



Fig. 3-1. Flip-flop arrangement of two triodes. Each successive negative 
pulse on the single input lead will reverse the operating state of the circuit. 
In the state shown (Vi conducting) a negative pulse lowers the cathode 
potential of V% below its grid potential, causing F 2 to start conducting, 
which in turn lowers the grid potential of Vi. 

second is connected to the grid of the first. An increase in 
current through either triode tends to lower the grid potential 
of the other triode. For example, an increase of the plate current 
of Vi causes a greater potential drop across the resistor Ri, this 
lowers the grid potential on V%. Accordingly, the plate current 
of tube V 2 is decreased, which, in turn, lowers the potential drop 
across the resistor R 2 , causing an increase in the grid potential 
of tube Vi. With a higher grid potential on tube Vi, the plate 
current through this tube is increased. This general condition 
of instability prevails and continues to drive the operation just 
described until tube Vi is conducting at saturation and tube V 2 
is cut off. The plate current of the conducting tube produces a 
voltage drop across both its own plate resistor and the common 
cathode resistor R3. The values of the resistors are selected so 
that the potential at both cathodes, with either tube fully con- 
ducting, is above the grid potential of the nonconducting tube. 
This elementary configuration can be caused to shift from one 



COUNTERS AS ELEMENTARY COMPONENTS 



15 



stable condition to the other by introducing circuitry designed 
to respond to various stimuli. There are two main types of 
shifting stimuli which can be introduced to the flip-flop circuit, 
and both of these have application in computing devices. 

The first type of stimulus utilizes pulses of a fixed polarity 
which are introduced at a single input terminal. The action of 
the flip-flop is such that it reverses between its two stable states 
each time a new input signal is introduced : it is similar in action 

NEON LAMPS 
M P 




-115 TO 
-70 V c 

INPUT— |f ' 



-85 



Fig. 3-2. Binary counting circuit with a single input lead. 

to a pull-chain light switch which reverses its state (on or off) 
each time the chain is pulled. Thus it indicates, by its state, 
whether the total number of input pulses which have been 
applied is odd or even. 

Such a device is classified as a modulo 2 counter. The term 
modulo means, literally, "with respect to the modulus, or stand- 
ard of measurement." Two numbers are congruent with respect 
to an integral modulus if their difference is divisible by that 
modulus. That is, a = b (mod c) if a — b is divisible by c (or 
if a and b have the same positive remainder when divided by c). 

A modulo 2 (or binary) counter is essentially the main ele- 
mentary component from which counters and accumulators are 
assembled. A practical example of such a binary ring circuit 
having a single input lead, as found in the plus-minus indicating 
device of the ENIAC, is illustrated in Fig. 3-2. In this illustra- 
tion the tubes V\ and Vi act as trigger tubes. Whenever a 
positive signal arrives on the one input lead, both of these tubes 



16 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

begin to conduct. This causes the grids and plates of both V 3 
and Vi to take a negative swing. Since F4 is already noncon- 
ducting, it has no effect there, but F 3 is cut off. As F 3 cuts off, 
its plate becomes more positive, and, through condenser C 3 , 
the grid of V ± also takes a positive swing. As V4, begins to con- 
duct, its plate takes a negative swing, and condenser C 2 further 
causes the tube F 3 to cut off. Note that the tubes utilized are 



100 k 2 

vwv — -»- (-435 v) 



25/^/ 




(-465V) 



.001 =t= 



+ 110 -555 -360 — s/WV 

I00K I00K 



is 

INPUT 
(SET) 



INPUT 
(RESET) 



Fig. 3-3. Binary counting circuit with two input leads. 



actually dual triodes so that Vi and V2, and likewise Vz and T 4 , 
are in a single envelope. 

The second type of actuating signal, or stimulus, for altering 
the stable state of the flip-flop is introduced on one of two input 
terminals. If the input signal arrives on the first terminal, it 
causes the flip-flop to assume the first of its two stable states and 
remain in this state thereafter until such time as a signal arrives 
at its other input terminal, whereupon it will reverse. This 
action is analogous to that of the ordinary wall-type toggle switch, 
which can be thrown in either of two directions and which will 
remain thrown in this direction until such time as it is thrown the 



COUNTERS AS ELEMENTARY COMPONENTS 17 

other way. A practical example of such a side-stable electronic 
circuit is shown in Fig. 3-3. Except for a few circuit values, the 
only difference between this and the circuit of Fig. 3-2 is in the 
number of input leads for introducing the actuating signal. 

3-4. Pulse -sensitive Flip-flop Circuits 

Of the two types of flip-flop circuits described above, the one 
having a single input terminal is of interest mainly because a 
whole series of pulse-sensitive counters can be built from it. 
When considered as a single unit operating by itself, it is essen- 
tially an odd-even, or modulo 2, counter, as mentioned previ- 
ously. If a number of such modulo 2 counters are connected 
so that the output of one feeds the input of the next, a radix 2 
counter is formed, each successive flip-flop registering the next 
binary digit. (The reader who is not familiar with the concepts 
of binary arithmetic is referred to Chap. 6 for an exposition which 
will explain more fully the meaning of this paragraph.) If the 
number of tubes in the flip-flop itself is increased beyond two, 
the circuit can count to a modulus higher than 2. Stating the 
proposition generally, a number of modulo r flip-flops used 
together constitute a counter system operating to the radix r. 
It is also possible to connect a number of flip-flop circuits in 
such a manner that each complete flip-flop represents one ele- 
ment in a ring of such twin elements, thus forming another type 
of modulo r counter, which can be compounded into a radix r 
counter. 

All these configurations are characterized by the presence 
of a single input terminal and further by the fact that they are all 
fundamentally compounded from a number of modulo 2 flip-flop 
elements. Such compounding is necessary because, although it 
is easy to think of all radices as having substantially equal 
mathematical merit, the electron tube is unable to assume more 
than two discrete stable states with any assurance of reliability. 

3-5. Radix 2 Counters 

One of the simplest schemes for counting the pulses in a series 
is the radix 2 counter, which may be assembled by coupling 
together several modulo 2 counters. 15 ' 24 ' 34,43 Figure 3-4 illus- 
trates how this may be done. The circuit shown contains four 



18 THE BASIC ELEMENTS OF MACHINE COMPUTATION 




.._ip- 



COUNTERS AS ELEMENTARY COMPONENTS 19 

modulo 2 counters denoted by the blocks with the Roman 
numerals. Counter I receives the pulses to be counted; these 
pulses are introduced via the input lead of a modulo 2 counter 
of the type illustrated in Fig. 3-2. Each pulse causes counter I 
to reverse its state. The output from counter I drives counter II. 
The output is taken from point B in Fig. 3-2 and is coupled, 
possibly through a pulse-sharpening circuit, to the input lead of 
counter II. In the same way the output from point B of counter 
II is taken to the input of counter III, and the output at point B 
of counter III is taken to the input of counter IV. Now it is 
clear that each time tube F 4 of counter I, II, or III changes 
from a conducting to a nonconducting condition, a positive pulse 
will be introduced to the input of the next counter in the series, 
causing it to change its state. Assume that a counter's condition 
is denoted by the symbol if tube Vi is blocked and tube V 3 is 
conducting. Then a pulse changing the counter's status from 
to 1 introduces a negative pulse which has no effect on the next 
counter in the series. However, if the status is changed from 1 
to 0, a positive pulse is introduced to the next counter, and this 
pulse changes the status of the next counter. Thus, in the 
example of Fig. 3-4, if the configuration of the four counters 
were 0000 at the beginning and if 16 pulses were introduced, the 
successive configurations would be 0001, 0010, 0011, 0100, 0101, 
0110, 0111, 1000, 1001, 1010, 1011, 1100, 1101, 1110, 1111, and 
0000. 

The representations 0000 through 1111 are the radix 2, or 
binary, representations of the decimal numbers from through 15. 
The 16th pulse returns the system to zero. This representation 
is described more fully in Chap. 6; it should be noted here, how- 
ever, that from the configuration of the system the number of 
pulses which have been introduced can be deduced, provided this 
number is less than 16. By adding more counters modulo 2 
with similar coupling, the capacity of the circuit can be 
increased. The number of pulses which can be counted with m 
modulo 2 counters is 2 m — 1. 

The reading of these counters is brought about without sub- 
stantial decrease in their functioning frequencies by establishing 
reading leads at points A of the individual modulo 2 counters. 
Simple radix 2 counters can be made to operate at a counting 
rate of 10 pulses per microsecond. 



20 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

3-6. The r-triode Counter, Modulo r 

For small values of r, such as 3, 4, and 5, it is feasible to extend 
the circuit of Fig. 3-1 beyond two, so that, of r triodes, one is 
conducting and r — 1 are blocked. 2,15>17,37 Thus, a pulse intro- 
duced to the input lead causes the originally conducting tube to 




pUU- 



Fig. 3-5. Five-triode ring. This circuit has five stable states, in each of 
which one, and only one, tube is conducting. It is a modulo 5 counter. 

become blocked and causes a tube cyclically advanced beyond 
the originally conducting tube to begin to conduct. Figure 3-5 
illustrates such a circuit 34 with five triodes so wired that the 
direction of progress is counterclockwise. In this circuit the 
grid resistors are chosen carefully so that the condition in which 
none, or more than one, of the tubes is conducting leads to a grid 
voltage too high or too low for stable operation. The grid of 
each tube obtains its operating potential from a voltage divider 
comprising five resistors, four from the plates of the four other 



COUNTERS AS ELEMENTARY COMPONENTS 21 

tubes, and one from a permanent bias bus. The values are 
such that, for any one tube, only two conditions can obtain: 
(1) if the other four are all cut off, the grid voltage will be suffi- 
ciently high that full conduction of this tube is assured, or (2) if 
one of the other four tubes is conducting, then the grid voltage 
will be low enough to assure that this tube is cut off, thus pre- 
serving the one-on-four-off ratio. 

A negative pulse on the input-pulse bus will cause the conduct- 
ing triode to cut off, thus raising its plate potential. This rise is 
then transmitted to the other four tubes. However, only the 
associated voltage-divider resistor of the next tube cyclically 
advanced is shunted by the necessary crossover condenser; 
hence it is the only one to achieve conduction as a result of the 
input pulse. [It is also possible 41 to construct a circuit such 
that only one tube is cut off and the rest are fully conducting. 
Another possibility 23 is a ring composed of an odd number of 
tubes, r, of which in each of the ring's r stable positions (r + l)/2 
tubes are nonconducting.] 

Read-out from the modulo r counter illustrated in Fig. 3-5 is 
accomplished at the plates of the tubes, as illustrated by the 
small neon indicator lamps. Its operation is limited to values of 
r which are sufficiently small to permit adequate discrimination 
between conditions governing operation of the individual tubes. 
In the configuration shown, the plate potential of the conducting 
tube may be well below that of the nonconducting tubes. The 
difference in potential between the plates and the grids of the 
nonconducting tubes is 1/r times the difference between the 
plate of the conducting tube and the plates of the nonconducting 
tubes. If r becomes sufficiently large, 1/r becomes sufficiently 
small to make operation of the circuit uncertain. In practice, 
with careful design, values of r up to 7 or 8 appear to be feasible ; 
values much larger than this have led to difficulty. It does not 
appear possible to make r = 10 (for decimal arithmetic) with 
this scheme, except at considerable decrease in reliability. 
However, a binary counter and a quinary counter can be com- 
bined to form another type of decimal counter. 15,23 - 34 

There is another form of r- triode modulo r counter which is 
practical and which has achieved wide usage in certain specialized 
applications where the ultimate in speed is not required. This 
is the ring consisting of gas-filled triodes, 15,40 - 42 ' 43 popularly 



22 THE BASIC ELEMENTS OF MACHINE COMPUTATION 



known as thyratrons. The thyratron, as contrasted with the 
conventional tube, has only two possible stable states for all 
grid potentials: fully conducting and completely nonconducting. 
The grid of such a tube, by itself, serves to trigger the tube from 
the nonconducting to the fully conducting state, after which it 
loses control of the action of the thyratron until the tube is 
extinguished by other means. Figure 3-6 represents a decade of 
thyratrons, forming a modulo 10 counter. In each of its 10 stable 
states, one, and only one, tube is conducting. Because of the 



INPUT 
+ PUL*§E 




♦ c- 



Fig. 3-6. Thyratron ring counter. 

voltage divider in the cathode circuit of the conducting tube, the 
grid voltage of the tube following it is higher than the grid 
voltages of the other nonconducting tubes. The tube is there- 
fore said to be primed. The next positive pulse raises its grid 
sufficiently to cause it to start conducting. Almost simul- 
taneously the preceding tube is cut off. This is the consequence 
of a drop in its plate voltage while its cathode voltage remains 
nearly constant. The drop in plate voltage is occasioned by the 
additional current (to the tube which has begun to conduct) 
through the common plate resistor; at the same time, the cathode 
potential is held nearly constant by the action of the capacitor 
in the cathode circuit. 

This gas-tube ring counter can be used in applications requiring 
a rate up to about 10,000 counts per second, this limit being 
established probably by the time constant of the grid circuit. 



COUNTERS AS ELEMENTARY COMPONENTS 23 

The ultimate limitation on counting speed is set by the deioniza- 
tion time of the gas molecules. It is probable that 50,000 counts 
per second is about the best that will be reliably accomplished 
with production gas-tube counter circuits. Higher speeds, of 
the order of 1 million counts per second, demand the use of 
vacuum tubes exclusively, although experimental gas-tube rings 
have been made to work as fast as 110,000 to 200,000 counts 
per second. 39 

3-7. The 2r-triode Counter, Modulo r 

For high-speed counting to a radix greater than 2, the 2r-triode 
counter modulo r unit is probably the only answer. The most 
advanced present form of this type of counter will be found in the 
decade rings of the ENIAC, where it forms the basic counting 
unit. An example of such an arrangement taken from the 
ENIAO 34 ' 39 is shown in Fig. 3-7. 

It will be noticed that there is a strong resemblance between 
the circuit of Fig. 3-7 and the more elementary circuit of Fig. 3-1. 
The former is merely a tenfold repetition of the latter. It is 
interesting to note, however, that this circuit can be operated 
more reliably if the coupling between flip-flop stages is effected 
through the regular control grids while the single input terminal 
is connected to the cathode circuits on only one side of each flip- 
flop. Thus the condition of the ring as a whole assures that 
there is one flip-flop in the "set" or abnormal state and that the 
other nine are in the "reset" or normal state. Successive input 
signals cause the flip-flop stage previously in the abnormal state 
to resume its normal condition, meanwhile simultaneously driving 
the next succeeding flip-flop into the abnormal state. 

As compared with the counter described in the previous section 
(i.e., the one with a single triode per digit of the modulus), the 
particular counter described above does not appear to have any 
practical limit as to the size of the modulus, although of course 
considerably more driving power may be required in the case of 
the larger rings. For computing purposes, 10 seems to be the 
highest modulus usually desired. 

It is possible to combine a number of modulo 10 counters in 
such a manner that a carry-over is provided between successive 
rings; this in effect gives us a radix 10, or decimal, counter 
capable of handling as many digits as there are rings provided. 



24 THE BASIC ELEMENTS OF MACHINE COMPUTATION 









-IQ. 

2? 







4\h 



*v 



in q: 



3 'C 

m ^ 

-° .a 
§ fl 

° O 
A « 

* a 

8.13 

-»J „ 

03 ! - 
"^ w 

« 3 

^3 

<3 O 

-8 S 



s a 



.a-s 

^ ° 

•3 e 
•SI 

fafi O 

a n 



COUNTERS AS ELEMENTARY COMPONENTS 



25 



Again there seems to be no limit to the number of rings that can 
be combined in this carry-over fashion. 

As a final example of this type of counter, we refer again to 
the ENIAC, which is capable of handling up to 20-digit decimal 
numbers. One accumulator in the ENIAC contains 10 rings of 
10 flip-flops each, or a total of 200 tubes, which are enclosed in 
100 envelopes. Actually each such 10-digit accumulator requires 
almost three times this number of tubes in order to handle certain 
auxiliary functions such as signal shaping, carry-over, add 
outputs, subtract outputs, and resets. 34,38 

The choice between the r-triode counter and the 2r-triode 
counter is completely dependent upon the value of r itself. In 
the ENIAC, for example, there are parts of the machine where 
r = 10 (all the decimal arithmetic), others where r = 6, 9, 11, 
13, 14, or 20 (various special-purpose rings), and several parts 
where r = 2 (plus-minus indicators). In all the cases where 
r > 2, experiments showed that the 2r-triode counter was 
required; for r = 2, however, the r-triode ring, or elementary 
flip-flop, was preferable and, in fact, was capable of operating 
about twice as fast as the 2r decade rings. 

3-8. Other Types of Flip-flop and EPDCC 

In addition to the foregoing general types of EPDCC's, there 
are a number of special ones which differ in engineering aspects 
and which have application in certain special circumstances. 
These will be summarized briefly below. 

There is one type of flip-flop which is particularly interesting 
in view of its simplicity and the small number of component parts 
required. This is the pentode-coupled flip-flop (Fig. 3-8), in 




50 K 



INPUT 

PULSES 

+ 8V. 



Fig. 3-8. Pentode-coupled flip-flop circuit. 



26 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

which the screens and plates of the two pentodes are cross- 
connected. 3,28 The action of this circuit can be readily under- 
stood by considering the case in which Vi is fully conducting; 
its plate must be at a low potential, and therefore the screen of 
V2 is at an equally low potential. This being the case, the right 
pentode must be essentially nonconducting, which means that 
its plate is at a high potential. Hence the screen of the left 
pentode, which must be at the same high potential, is serving to 
keep the left pentode in its fully conducting state. Positive 
input impulses serve to switch this flip-flop alternately into its 
two stable states, just as in the case of the triode flip-flop in 
Fig. 3-2. This circuit does not appear to have had the exploita- 
tion and development which its simplicity warrants. It operates 
satisfactorily at a speed of at least 200,000 counts per second; 
however, it appears likely that with an appropriate amount of 
effort directed toward its fullest exploitation this circuit is capable 
of the same speeds as are now achieved in the more fully developed 
triode flip-flop. It is quite possible that the advent of printed- 
circuit techniques, 5 mentioned in Sec. 16-6 of this book, and the 
recently released wartime developments of miniaturized vacuum 
tubes may serve to direct further interest toward the pentode- 
flip-flop circuit. 

The pentode-flip-flop circuit has been incorporated into a 
decade counter. 29 This particular decade differs from others 
previously described in that there are only five pairs of pentodes 
constituting the entire decade ring; thus, we might classify this 
as an r-pentode counter, modulo r (with the proviso that r be an 
even number). This particular circuit is also interesting in that 
its normal mode of operation is one in which there are r/2 con- 
secutive pentodes in the conducting state and r/2 consecutive 
pentodes in the cutoff state. (The two pentodes forming any 
one pair are r/2 apart in the ring.) As opposed to the more 
conventional counter ring, which has one element conducting 
and r — 1 elements nonconducting, this may appear to com- 
plicate the read-out problem; however, a rather ingenious and 
unambiguous visual read-out system has been developed for 
this pentode ring, and doubtless the same principle could be 
applied to the read-in and read-out problems which would arise if 
such a ring were integrated into a complete computing machine. 
This five-pentode-pair ring has been built and tested up to 



COUNTERS AS ELEMENTARY COMPONENTS 27 

100,000 counts per second or more. It compares favorably with 
the performance of the standard production models of triode-pair 
decade counters (180,000 pulses per second), which were brought 
to an advanced state of development for the ENIAC. More 
recently this same circuit has been further developed, 33 and it now 
appears capable of achieving a speed of 500,000 counts per second. 

It will be noted that all the foregoing decade rings operate 
with a fixed direction for the progression of the count. This is 
true for all the rings (r-triode, 2r-triode, r-pentode) except, of 
course, for the trivial case of the binary counter. In general, 
this is quite satisfactory for any type of computing machine, 
since the forward direction can be made to serve for addition, 
while a system of complements to the forward direction can be 
made to serve for subtractive purposes. It is a relatively simple 
problem to provide a few extra tubes for such a system of numbers 
and complements; this is actually the method employed in the 
ENIAC. 38 However, it is possible to fabricate a ring which can 
be driven either forward or backward. At least one such 
device has actually been constructed; the choice of drive direction 
in this ring was determined by the choice of which of two pairs 
of input terminals received the input pulses. 30 Since these 
flip-flops and counter rings are now in a state of rapid develop- 
ment, it is not possible to include here all the numerous finer 
engineering points. 

One more type of electronic counter should be included here, 
although its use to date has been directed more toward a simple 
counting device than as a component in a full-scale computer. 
This is the decade scaler, or decade counter, which counts in the 
decimal system, but not in a manner similar to previously 
described counters (all of which embodied a closed ring or 
chain of 10 electronic elements). This counter actually employs 
a chain of four binary-system counters similar to the ones 
described earlier in this chapter. These would, of course, nor- 
mally count on a modulus of 2 4 = 16, but certain trick feedback 
circuits are employed so that effectively six of the 16 stable 
states are circumvented in one way or another, thus reducing 
each chain of four flip-flops to what is actually a chain of decimal 
counters. This type of circuit has been known for several 
years ; 4,18,26 ' 41 several manufacturers have brought out com- 
mercial versions of these circuits. The principal merit of this 



28 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

modified-binary type of decimal counter as compared to the 
more conventional ring of 10 lies in the lowering of the input 
electrostatic capacity, which permits the circuit to achieve a 
higher operating speed. There is also considerable tube economy 
as compared to the 2r-triode ring when r becomes 10 or more. 

Two other types of flip-flop might be included here for the sake 
of completeness and because of the theoretical interest they hold. 
Neither of them has to date achieved any acceptance in practical 
computer application. The first of these is a standard cathode- 
ray tube modified in such a way that the beam, instead of falling 
on a phosphorescent screen, will fall on one of two collector 
plates, where it will be picked up and used to control the deflec- 
tion of the beam so that it continues to remain on this first col- 
lector plate. 39 By suitably arranging the geometry of the 
various internal members and the external circuitry of the tube, 
a satisfactory flip-flop element has been obtained. If there 
were enough demand for such a tube, it would not be too difficult 
to effect the necessary design and development work; such a tube, 
however, does not appear to offer any great advantage over 
presently available components, and there appear to be a number 
of engineering disadvantages to such a scheme. 

Another type of flip-flop may be built around the use of any 
single tube which shows an inversion in its characteristic curve. 
When such an inversion exists, it may be used to represent, 
effectively, an area of negative resistance, and, with the proper 
choice of operating conditions, this may be used to obtain the 
necessary two stable states. 15 However, these regions of nega- 
tive resistance are caused by phenomena such as secondary 
emission, which is notoriously unstable and unreliable. This 
type of tube does not appear to promise any major gain over 
present conventional components. 

3-9. Economic Considerations 

Experience with the cost of existing and contemplated machines 
(described in Chap. 10) does not serve as a very accurate guide 
toward establishing a cost of the elementary components, because 
the cost of the few existing equipments is intimately associated 
with developmental charges and certain military considerations. 
For this reason it appears best to look to other sources of data. 
For the purposes of this estimate, we shall exclude EPDCC's 
of the elementary mechanical type. The cost figures for desk 



COUNTERS AS ELEMENTARY COMPONENTS 29 

computers given in Chap. 8 may be taken as indicative of what 
mechanical devices are likely to cost. 

Manufacturers are presently engaged in making standard 
models of electronic counting equipment on a limited scale. 
The Radio Corporation of America is now producing a six-decade 
decimal counter (operating on the modified-binary system 
described in Sec. 3-9) which is capable of operating on pulses 
spaced 1 million per second. 27 This device is supplied with an 
input for normal pulse counting or with a start-stop input control 
which operates a self-contained, 1-megacycle, crystal-controlled 
oscillator. The Potter Instrument Company is also producing a 
Decade Scaler having only three decades (also modified-binary 
type) but capable of a speed of only 200,000 pulses per second. 25 

In general, it may be stated that the cost of an elementary 
flip-flop or a decade ring is not closely governed by the operating 
speeds involved. This statement is approximately true over 
quite a wide range of speeds: from about 100 cycles per second, 
the limit of mechanical counters, to several million cycles per 
second, the present limit of electronic counters. Speed require- 
ments may dictate the choice of tubes involved and their asso- 
ciated circuitry. So far as flip-flops and rings are concerned, the 
choice of tubes and circuitry does not greatly influence the cost 
of the over-all assembly, since labor appears to be by far the 
greatest factor. For example, a top-speed modified-binary 
decade employing four flip-flops (eight tubes plus a few extra 
tubes) will have about as many tubes as a low-speed thyratron 
ring of the conventional type. The two may, therefore, be 
regarded as about equally expensive even though their operating 
speeds may differ by a factor of over 100 to 1. 

Labor is the principal item of expense per EPDCC. The use of 
recently announced laborsaving techniques such as the printed 
circuit may be expected eventually to reduce this; for example, 
printed circuits have been accepted in fabrication of radios and 
hearing aids. 5 Their potentialities have not yet been exploited 
in connection with high-speed computers. It is not unreasonable 
to suppose that eventually the quantity cost per EPDCC may 
be brought down to the neighborhood of $1. 

REFERENCES 

1. Aiken, H. H., and G. M. Hopper, The Automatic Sequence Controlled 
Calculator — I, Electrical Engineering, Vol. 65, Nos. 8 and 9, pp. 384-391 



30 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

(August-September, 1946); The Automatic Sequence Controlled Calcu- 
lator— II, ibid., No. 10, pp. 449-454 (October, 1946); The Automatic 
Sequence Controlled Calculator — III, ibid., No. 11, pp. 522-528 (Novem- 
ber, 1946). 

2. Alfven, H., A Simple Scale-of-Two Counter, Proceedings of the Physical 
Society (London), Vol. 50, pp. 358-359 (May, 1938). 

3. Blair, J. M., An Improved Current Integrator, Review of Scientific 
Instruments, Vol. 14, pp. 64-67 (March, 1943). 

4. Blume, Richard J., Predetermined Counter for Process Control, Elec- 
tronics, Vol. 21, No. 2, pp. 88-93 (February, 1948). 

5. Brunetti, Cledo, and R. W. Curtis, Printed Circuit Techniques, Pro- 
ceedings of the Institute of Radio Engineers, Vol. 36, pp. 121-161 (Jan- 
uary, 1948). 

6. Burks, A. W., Electronic Computing Circuits of the ENIAC, Proceed- 
ings of the Institute of Radio Engineers, Vol. 35, pp. 756-767 (August, 
1947). 

7. Compton, J. S., U.S. Patent 2,402,372; filed Nov. 5, 1942; issued Sept. 
10, 1946. 

8. Couffignal, P. L., U.S. Patent 2,318,591; filed March 26, 1937; issued 
May 11, 1943. 

9. Crawford, Perry, Jr., unpublished thesis, Massachusetts Institute of 
Technology (1939). 

10. Desch, J. R., U.S. Patent 2,404,697; filed March 21, 1942; issued July 
23, 1946. 

11. Eccles, W. H., and F. W. Jordan, A Trigger Relay, Radio Review, Vol. 1, 
pp. 143-146 (October, 1919). 

12. Grosdorf, I. E., Electronic Counters, RCA Review, Vol. VIII, pp. 438- 
447 (September, 1946). 

13. Harvard University, Staff of the Computation Laboratory, A Manual of 
Operation for the Automatic Sequence Controlled Calculator, Annals of the 
Computation Laboratory of Harvard University, Vol. I (Harvard Univer- 
sity Press, Cambridge, Mass., 1946). 

14. International Business Machines Corporation, IBM Automatic Sequence 
Controlled Calculator. Pamphlet published by the International Business 
Machines Corporation (1945). 

15. Lewis, W. B., Electrical Counting (Cambridge University Press, New 
York, 1942), Chap. VIII. 

16. Lifschutz, H., and J. L. Lawson, Triode Valve Scale-of-Two Circuit, 
Review of Scientific Instruments, Vol. 9, pp. 83-89 (March, 1938). 

17. Lifschutz, H., A Complete Geiger-Miiller Counting System, Review of 
Scientific Instruments, Vol. 10, pp. 21-26 (January, 1939). 

18. Lifschutz, H., New Vacuum Tube Scaling Circuits of Arbitrary Integral 
or Fractional Scaling Ratio, Physical Review, Vol. 57, pp. 243-244 (1940). 

19. Michael, F. R., Tube Failures in ENIAC, Electronics, Vol. 20, No. 10, 
pp. 116-119 (October, 1947). 

20. Miller, O. R., U.S. Patent 2,407,320; filed Nov. 5, 1942; issued Sept. 10, 
1946. 

21. Neher, H. V., High Speed Mechanical Recorder, Review of Scientific 
Instruments, Vol. 10, pp. 29-31 (1939). 



COUNTERS AS ELEMENTARY COMPONENTS 31 

22. Overbeck, W. P., U.S. Patent 2,404,918; filed May 1, 1940; issued 
July 30, 1946. 

23. Overbeck, W. P., U.S. Patent 2,470,716; filed June 11, 1943; issued 
May 17, 1949. 

24. Phelps, B. E., Dual-Triode Trigger Circuits, Electronics, Vol. 18, 
No. 7, pp. 110-113 (July, 1945). 

25. Potter Instrument Company, Flushing, N. Y., Decade Scaler, Model 2092. 

26. Potter, J. T., A Four-Tube Counter Decade, Electronics, Vol. 17, No. 6, 
pp. 110-113, 358, 360 (June, 1944). 

27. Radio Corporation of America, Camden, N.J., Time Interval Counter, 
Type WF-99B. 

28. Regener, V. H., Design and Use of Directly Coupled Pentode Trigger 
Pairs, Review of Scientific Instruments, Vol. 17, pp. 180-184 (May, 1946). 

29. Regener, V. H., Decade Counting Circuits, Review of Scientific Instru- 
ments, Vol. 17, pp. 185-189 (May, 1946). 

30. Regener, V. H., Reversible Decade Counting Circuit, Review of Scientific 
Instruments, Vol. 17, pp. 375-376 (October, 1946). 

31. Reich, H. J., Trigger Circuits, Electronics, Vol. 12, No. 8, pp. 14-17 
(August, 1939). 

32. Reich, H. J., Electronic Switch for Simultaneous Observation of 2 
Waves with the Cathode-ray Oscilloscope, Review of Scientific Instru- 
ments, Vol. 12, pp. 191-192 (April, 1941). 

33. Seren, Leo, Decade-Ring Scaling Circuit, Review of Scientific Instru- 
ments, Vol. 18, pp. 654-659 (September, 1947). 

34. Sharpless, T. K., High Speed N-Scale Counters, Electronics, Vol. 21, 
No. 3, pp. 122-125 (March, 1948). 

35. Shea, H. G., Electronic True Decade Counters, Electronic Industries, 
pp. 82-84, 136 (Sept. 1946). 

36. Shepherd, W. G., and R. O. Haxby, Scale of Eight Impulse Counters, 
Review of Scientific Instruments, Vol. 7, pp. 425-426 (1936). 

37. Stevenson, E. C, and I. A. Getting, A Vacuum Tube Circuit for Scaling 
Down Counting Rates, Review of Scientific Instruments, Vol. 8, pp. 
414-416 (November, 1937). 

38. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, The EN I AC (University of Pennsylvania, Philadelphia, Pa., 
June, 1946). 5 vols. 

39. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, Theory and Techniques for Design of Electronic Digital Com- 
puters (University of Pennsylvania, Philadelphia, Pa., 1946). 2 vols. 

40. Wellman, B., and K. Roeder, Electronic Counter for Rapid Impulses, 
Electronics, Vol. 15, No. 10, pp. 74, 140-142 (October, 1942). 

41. Wild, John J., Predetermined Counters, Electronics, Vol. 20, No. 3, pp. 
121-123 (March, 1947). 

42. Wynn- Williams, C. E., The Use of Thyratrons for High Speed-Auto- 
matic Counting of Physical Phenomena, Proceedings of the Royal Society 
{London), Vol. 132, pp. 295-310 (1931). 

43. Wynn- Williams, C. E., Thyratron "Scale-of-Two" Automatic Counter, 
Proceedings of the Royal Society {London), Vol. 136, pp. 312-324 (May, 
1932). 



CHAPTER 4 
SWITCHES AND GATES 
4-1. Introduction 

The basic operations of any digital computer are the ordinary 
elementary arithmetic operations, addition, subtraction, multipli- 
cation, and division, all of which can be performed on most of the 
inexpensive desk computers as explained in Chap. 8. A principal 
feature which distinguishes the organized digital computer 
(described in Chap. 5) from a mere aggregation of desk com- 
puters is the ability to switch data rapidly from one component 
of the computer to another without the necessity of transcribing 
the intermediate results to paper either by a hand process or by a 
printing procedure. Other features are the speed with which the 
individual operations are performed and the extent of the internal 
storage. The most distinguishing characteristic, however, is 
the ability to switch numbers, representing either intermediate 
solutions or coded instructions, from one arithmetic or control 
unit to another, between the arithmetic or control units and 
storage, and between the storage and input or output equipment. 
This chapter describes the various uses of switching equipment, 
the various types of switching equipment which are presently 
available or which may be developed and exploited in the near 
future, and examples of the practical application of these switches 
to portions of a complete digital computer. 

In the subsequent discussion of such equipment, a much-used 
type of switch is denoted as a gate. This is an on-off switch 
in which the passage of one electrical signal is controlled by the 
presence of one or more other signals which hold the switch on 
or off. 

4-2. Switch Functions in Computers 

Switches and gating equipment are used in many places 
throughout a digital computer, and, in fact, the same type of 
equipment may be used in a number of different places in a 
computer for widely different purposes. Each application of 

32 



SWITCHES AND GATES 33 

such a switch or gate, however, falls into one of three general 
categories outlined below. 

4-2-1. Data Switching. The Automatic Sequence Controlled 
Calculator, or the Harvard Mark I machine, offers what is 
probably the most easily visualized example of the switching of 
data from one part of a machine to another. 1 - 5 Briefly stated, 
this device consists of a number of components including storage 
registers, constant registers, and multiply-divide units, each of 
which has both an input and an output circuit and a single 
number-transfer bus. Each elementary partial step in a com- 
putation is carried out by connecting the output of one unit and 
the input of another unit to the number-transfer bus. The result 
attained is the transfer of the contents of one unit to the other 
unit by means of the number-transfer bus. At any instant, 
there is at most one output circuit and at most one input circuit 
connected to the number-transfer bus. After this arithmetic 
operation has been completed, another pair of registers may be 
connected to the common number-transfer bus. The choice of 
which units shall be so connected is made through other elements 
in the machine (such as the main sequence tapes, which are more 
fully described in Chap. 10). The actual means of carrying out 
the connections, disconnections, and reconnections is to be found 
in a large bank of electromechanical relays. 

Much the same sort of data switching may be found in other 
types of computing machinery. In the ENIAC, for example, 
instead of just one such bus there are several number-transfer 
busses, so that elementary operations previously described may 
be carried out simultaneously in various parts of the machine. 2,11 
However, the net effect is that each individual number-transfer 
bus may have connected to it at any instant at most one output 
circuit and at most one input circuit. In the EDVAC there is 
again a single number-transfer bus. 12 

4-2-2. Data Conversion. It frequently happens in computers 
that data must be converted from one representation to another. 
An example is the ENIAC, in which data are sometimes stored 
statically in vacuum-tube circuits and transmitted serially in 
trains of pulses. The conversion is one in which a given number 
represented on the ring counter (the digit 7 represented by the 
seventh flip-flop in the ring, for example) is transformed into a 
train of (seven) pulses for transmission to some other part of the 
equipment, such as another ring. 



34 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

4-2-3. Signal Preservation. A large number of high-speed 
switches or gates are employed in computing machines for 
purposes which have nothing to do with the actual arithmetic 
operations; they are required solely for signal preservation by 
pulse shaping. This application of gating circuits is to be found 
only in high-speed computers; such devices as the Mark I and 
Mark II computers and the Bell Telephone Laboratories and 
IBM relay computers are limited by physical inertia rather than 
electrical signal distortion. 

Since discrete values are represented by the presence or 
absence of pulses in various parts of digital computers, it is 
obvious that the identity of the individual pulses must be 
properly preserved. In a conventionally designed machine, 
the pulses are of proper size, shape, and timing at the time and 
place originated. However, after traveling through a long cable 
for transfer to some other part of the computer, or after being 
operated upon in a device such as an adder, or after being 
temporarily stored in some memory equipment, these pulses 
may undergo considerable deterioration unless proper steps are 
taken to reestablish their original form. The pulses may not 
only be diminished in amplitude (a situation corrected by the 
insertion of an amplifier), but also may be delayed in time of 
arrival or spread out on the time scale to such an extent that they 
may give a false representation of the digital value involved. 
Electronic gates offer the means for overcoming this undesirable 
situation. A gate may be used to allow a standardized master 
pulse, which is continuously generated and available throughout 
the computer, to be combined with the distorted signal in such 
a way that a properly timed and shaped portion of the distorted 
signal is selected by the master signal, and this selected portion 
is then passed on to the rest of the equipment. This type of 
signal preservation or pulse standardization is required in all 
machines which use pulses spaced 10 microseconds apart or less. 
In order to achieve reliable operation with a high-speed digital 
computer, this pulse standardization is inserted at frequent 
locations throughout the machine. 

4-3. Types of Switches 

Four types of switches, including vacuum-tube gates, are 
presently used in computers. These types vary widely in cost 



SWITCHES AND GATES 35 

and speed of operation ; each is suitable for one or more particular 
purposes. 

4-3-1. Nonautomatic Switching. Nonautomatic switching is 
characterized by cable connections and electrical switch settings 
which are manually adjustable; they are set up initially for a 
problem and are not altered during its run. Only the simplicity 
of this type of switching allows its inclusion here since it is 
incapable of being operated by any other part of a computer. 
A number of such plug connectors and hand-set switches are 
used for the insertion of constants in the Mark I and ENIAC. 
In the ENIAC there are a number of plug connectors for setting 
up the operations of a problem. Such preset switches and cables 
constitute a form of storage of the lowest order. In the ENIAC 
the various units are permanently wired to do the specific opera- 
tions of addition, subtraction, multiplication, division, taking 
the square root, and looking up function values.* The particular 
operations which these units perform in a given problem and the 
order in which they do them depend to a great extent upon how 
units are interconnected by the various cables and how the 
manual switches are set. With a machine of this type, the setup 
time is so long that it can be operated profitably only if it is 
called upon to handle a large number of similar problems which 
can all be run on a single manual setup. This is true of the type 
of problem which the ENIAC was designed to handle. 

4-3-2. Electromechanical Relays. The electromechanical re- 
lay exists in a wide variety of sizes and shapes. It is essentially a 
metallic switch, or several switches, which can be operated by an 
electrical signal. A relay includes an electromagnet which 
receives an incoming signal to operate or not operate. When the 
signal to operate is received, the coil is energized, and a nearby 
piece of magnetizable material called the armature is attracted 
and moves toward the coil. This motion, by means of a mechan- 
ical linkage, causes an assembly of contacts to open or close. 

There is almost no limit to the arrangements possible with 
these contacts. For example, the contacts of a single relay may 
be made so that in one operation they will close certain circuits, 
open others, or transfer a circuit from one place to another. The 

* This statement is true of the machine as originally used. The philos- 
ophy of programming the machine has been altered since then, as 
explained in Chap. 10, to make the machine more versatile. 



36 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

removal of the input signal from a simple relay will deenergize 
the coil, and a spring will cause the contacts to return to the 
original position. Some relays are provided with mechanical 
or electrical latches which effectively maintain the contact in the 
operating position until a reset stimulus arrives through a differ- 
ent circuit. There are other relays, popularly called stepping 
switches, which must receive a series of alternate operate and 
nonoperate stimuli before returning to the original state. These, 
too, have limited application in computing equipment. For 
most computing purposes, however, the most useful type of 
relay is one which is capable of closing a number of electrical 
circuits when operated and reopening them when the operate 
signal is removed. Small relays have been designed for closing 
or for transferring 6 to 12 independent electrical circuits, and 
these are used effectively in the Mark I and Mark II machines. 

Since the operation of this type of relay is mechanical, a 
considerable length of time must elapse from the instant the relay 
coil is energized until contact is made, even though the distance 
traveled may be only a few hundredths of an inch. Six- and 
12-contact relays of the type mentioned are capable of operating 
reliably in about 15 milliseconds. Recently a number of relays 
capable of operating in 1 to 2 milliseconds have been announced 
by a relay manufacturer, 10 but these high-speed mechanical 
relays tend to be limited to one or two circuits. The wire-brush 
relay used by the IBM Company has an operating time of 
3 milliseconds and can operate eight contacts. The speed limita- 
tion of an electromechanical relay is a fundamental one, and it 
does not appear that an inexpensive mechanically operating 
relay can be produced which will overcome this limitation. 

Despite the limited speed of the mechanical relay, there are 
a number of significant advantages to this device. First, the 
relay is nearly a perfect switch; i.e., when the contacts are closed, 
they offer practically zero resistance to the passage of current, 
and when they are open, they offer an extremely high resistance. 
Second, as compared to electronic equipment of equal cost, the 
relays are capable of handling far more current. Also, one relay 
can handle 10 or more contacts. Therefore, a relay is attractive 
to a circuit designer because of its economy, relatively small size, 
and the large number of combinations of circuits which can be 
switched with a single relay. 



SWITCHES AND GATES 37 

Some relays are used in electronic computers in locations 
where speed is of lesser importance, e.g., in the reading of input 
information and its transfer to internal storage, or in the removal 
of data from internal storage to an output printer. There are 
numerous noncomputational requirements for relays in com- 
puters, such as in the control of power supplies, ventilators, 
etc. 

4-3-3. Vacuum-tube Gating Circuits. An electronic gate is a 
circuit with a single output and two (or more) inputs so designed 
that an output signal is produced when, and only when, input 
signals are received on both (or on a particular set of) input leads. 
Such circuits are variously known as gates, coincidence circuits, 
Rossi circuits, or logical and circuits. 

Figures 4- la, 6, and c illustrate three types of gating circuits. 
All perform the same operations and differ only in structural 
details and the polarity of signal required to operate them. All 
three of these circuits operate with nearly equal speed and are 
adequate for machines now in existence or presently contemplated. 



-•output c 

INPUT A •- 



inku I A • 1| m - • 

+ PULSE S 

I 



INPUT B • — |h»* 




Fig. 4-la. Dual grid gate. 

The simplest electronic gate employs a single tube which has 
two or more input grids. Tubes such as the 6SA7 and 6L7 are 
widely used in radio communications for frequency mixing; 
they may be used in a circuit such as shown in Fig. 4-la to form 
an effective and reliable gate. Except when input signals are 
applied, the two input grids are negative with respect to the 
cathode and the tube is therefore cut off. If either grid is driven 
positive, the condition remains substantially unaltered. How- 
ever, if both grids are simultaneously driven positive, the full 
electron current begins to flow, and the plate takes a large nega- 



38 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

tive swing which appears as an output signal. (The similarity 
between this circuit and the pentode-coupled flip-flop, Fig. 3-8, in 
which plate current could flow only when screen and control grids 
were simultaneously positive, should be noted.) One advantage 
of this circuit is that only a single tube is required. This single- 
tube circuit has the disadvantage that the two grids of conven- 
tional tubes do not have equal cutoff potentials; hence the two 
input grids are not strictly interchangeable. 

Figure 4-16 illustrates a Rossi circuit modified so that it is 
suitable for computing purposes. 8 In this circuit there are two 
tubes, possibly in the same envelope, the plates of which are tied 



B + 




^ output c 

~* +PULSE 



INPUT 9 
-PULSE 



Fig. 4-16. Rossi circuit, or parallel gate. 

together and fed through a single plate resistor. Both grids are 
normally positive with respect to the cathodes; the tubes are 
therefore conducting (as shown). Circuit parameters are so 
chosen that either of the tubes is capable of drawing all the plate 
current supplied by the batteries and plate-resistor combination. 
Thus, if one tube receives a negative input signal, the output 
voltage at the common plate connection rises very slightly. 
However, if both input circuits simultaneously receive negative 
pulses, then neither will draw current, and the output voltage 
will rise abruptly to the full plate battery potential. This is the 
action desired. The input stimuli are negative pulses, and the 
output (upon coincidence) is a positive voltage rise, in con- 
tradistinction to the multigrid-tube coincidence circuit previously 
described. The necessity for proper polarization of input signals 
is characteristic of all electronic switching equipment. It is 
obvious that a large number of switches may not be cascaded 
unless proper attention is paid to the polarity of the input pulses 



SWITCHES AND GATES 39 

and the resulting polarity of output pulses which are to be used 
elsewhere as inputs. 

In computing machines the two-stage circuit illustrated is 
generally adequate for most purposes. However, it is possible to 
extend this multitube coincidence circuit further; coincidence 
circuits with as many as 100 input stages have been successfully 
built and operated. Where only two input circuits are required, 
it is possible to use two simple triodes, which are readily available 
in a single envelope. For coincidence circuits involving a large 
number of input stages or for those which demand the ultimate 
in operating speeds, the more elaborate pentode is preferred. 

A circuit which apparently has not been used to date in com- 
puting machines, but one which merits consideration in the 



OUTPUT 
-PULSE 



Fig. 4-lc. Common cathode gate. 

design of future equipments, is shown in Fig. 4-lc. This is a 
combination of the gating circuit with the cathode follower.* 
This circuit is similar to Fig. 4-16 except that the common load 
resistance is changed from the plate to the cathode circuit. In 
the normal condition, both tubes are conducting. The values of 
the common cathode resistor and the bias-battery voltage are 
such that, with the tubes conducting, the cathodes are positive 
with respect to ground. If the input stimulus in the form of a 
negative voltage arrives at one grid, it will serve to cut off this 
one tube. As in the previous case, this will have only a trivial 
effect on the potential of the common cathodes or output circuit. 
However, if negative signals arrive on both grids, then both tubes 
are cut off, cathode current ceases to flow, and the output-circuit 
potential drops. An important feature of this circuit is that the 
polarity of the output signal is the same as the polarity of the 
input signals, thus providing the possibility of cascading a 



40 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

number of such circuits. This gate has the important advantage 
of any cathode-follower circuit in that it presents a considerably 
lower output impedance to the next circuit. This is an important 
consideration if the next unit in the computer is at a considerable 
distance and must be connected by a length of cable with a 
resulting high electrostatic capacity. This circuit and also the 
one of Fig. 4-16 can have identical tubes at each input and hence 
can be made with input leads which are interchangeable. 




OUTPUT 
C -PULSE 



INPUT 
PULSE 



+E VOLT PULSES]^ 4 
R. 



Fig. 4-2a. Resistance grid gate. 

4-3-4. Resistor Matrices and Diode Matrices. Attempts have 
frequently been made to simplify the electronic gating circuit by 
using pulses of equal size in a matrix of resistors or nonlinear 
elements such as crystal rectifiers. An elementary example of a 
resistor gate is shown in Fig. 4-2a. In this circuit the pulses 
must first be rigorously standardized as to size. In the example 
shown, the presence of a pulse is indicated by a potential of 20 
volts. The common junction point which feeds the grid circuit 
of the output tube may thus assume a potential of 0, +10, or +20 
volts, depending on whether zero, one, or two pulses arrive on the 
two input circuits. If the cathode is maintained at a potential of 
+ 15 volts by the batteries shown in the elementary configuration, 
the relative grid-cathode potential will be negative for the cases 
of zero or one input signal and positive only when two input 
signals are simultaneously present. In the latter case, plate 
current is permitted to flow and produces an output signal in the 
form of a negative pulse. By properly rearranging the circuit 
parameters, it is possible to cause this circuit to operate in the 
opposite fashion so that the coincidence of negative input pulses 
will produce an output pulse of positive polarity. This circuit 
works dependably with two input points and reasonably well 



SWITCHES AND GATES 41 

for slightly larger numbers. With a number of input circuits n, 
the threshold between operation and nonoperation is only the 
difference between (n — l)/n and 1. This precludes its use for 
values of n much larger than 5. Precisely the same limitation 
was found to exist in the r-triode counter, modulo r, described in 
the previous chapter. Furthermore, the circuit described fails to 
afford any significant economy over the ones illustrated in Figs. 
4-la, b, and c, because of the number of accessory tubes required 
to standardize the amplitude of the input signals. 

A considerable improvement in matrices can be obtained by the 
use of miniature crystal rectifiers, 3 ' 7,9 which were brought to a 



OUTPUT 

^~~ c - PULSE 




§£^-+ R 2» R i 

CURRENT FLOW CONVENTION 

Fig. 4-26. Diode-coupled grid gate. 

state of perfection for radar application during the recent war. 
Each of these diodes or crystal rectifiers contains a small piece 
of some semiconductor, such as germanium or silicon, with which 
contact is made by a small sharpened point, or cat whisker. 
Such a device presents a unilateral impedance. It is capable of 
carrying a small current in one direction while resisting up to a 
limit the passage of current in the reverse direction. A circuit 
similar to Fig. 4-2a is shown in Fig. 4-26 in which the pulse 
standardizers and resistor pair are replaced by two of these 
crystal rectifiers. A negative potential on either of the two 
crystals will hold the grid of the output tube below cutoff. 
However, if both crystals are simultaneously driven positive, 
they effectively disconnect their input circuits from the common 
connection point to the grid, thereby causing the potential at 
this point to rise, allowing the output circuit to conduct. Here, 
again, the circuit parameters may be rearranged so that negative 



42 THE BASIC ELEMENTS OF MACHINE COMPUTATION 



input signals are used and a positive output signal is obtained. 
The main improvement afforded by the crystals is the snap 
action of the potential change at the common tie point. This 
common point assumes one potential for both zero and one input 
signal and changes to its two-input potential only when two 
input signals are received. 

It is possible to expand this circuit to a large number of inputs 
since the snap action is obtained only when all inputs are driven 



A 



ION 
|OFF 



AX. 



ION 
fOFF 



— BINARY 
KEYBOARD 



A. 



A 



AXXX. 



8+ 



A. 



XX 



A 



AX. 



x. 



A 



AAXX 



A. 



AA. 



2° = i 



2' = 2 



2 Z =4 



7- 



OCTAL 



OUTPUT 

Fig. 4-3a. Binary to octal converter. For any one of the eight possible 
combinations of three on-off switch positions one, and only one, of the eight 
output terminals will have a positive voltage ; each of the other seven will be 
grounded through one or more of the diodes. 

in the proper direction. The main disadvantage of such crystal 
circuits is the relatively small amount of power which they are 
capable of handling. In practical application, an amplifier tube 
will usually be required to restore the power level after each 
diode array. However, for some applications with a number of 
diode switches at the same location, it is possible to defer this 
amplification until after several stages of crystal diodes. 

An example of such a situation is shown in Fig. 4-3a. This 
diagram represents a hypothetical converter which could be used 
for translating a number from binary to octal notation. The 
number 7 is shown on the binary keyboard on the left-hand side 



SWITCHES AND GATES 



43 



of Fig. 4-3a as 2 2 + 2 1 + 2°. The row of terminals labeled 1 to 7 
at the bottom of the diagram represents those digits in the octal 
system. It can be seen that every one of these terminals, except 
7, is grounded. The seventh terminal is the only one on which 
the voltage from the positive battery supply will appear. 

Figure 4-36 illustrates a translator which operates in the 
opposite direction. The input switch is set on number 7. 



b+. 



X. 



£jL. 



X 



X 



j^xxx 



x 



^x: 



X 



xx 



x 



X 



xxxx^ 



X 



XX 




ON 



OFF 



ON 



OFF 



2° = I 



2'= 2 



ON 



OFF 



2 Z =4 

BINARY 
OUTPUT 



Fig. 4-36. Octal to binary converter. For each one of the eight possible 
switch positions a different combination of output leads will be energized. 

It can be seen that all the on leads for the binary output are 
energized. It can also be seen that if the switch were on any 
of the other octal numbers, through 6, one or more of the on 
leads corresponding to some power of 2 would be grounded 
through the octal input switch. 

There is some inherent limitation to the rate at which pulses 
can be passed through a diode matrix; this is due primarily to 
the time constants of the diodes themselves. For computing 
applications, however, this limitation need not be considered a 
material design factor. 



44 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

4-4. Inverse Gates, or Buffers 

The gating circuits discussed in a previous paragraph are 
frequently described as being the electronic expression of a logical 
and; i.e., the output terminal C produces a signal only when the 
two input terminals A and B receive signals simultaneously. 
In computer design there is frequently a need for a circuit which 
functions as a logical or, to supply an output if either A or B is 
signaled. These logical or circuits are referred to as anti-Rossi 
circuits, isolating circuits, or simply buffers. The term logical 
or is used in the sense that A or B includes A and B. 

The difference between such a buffer and the gates previously 
described is simply the interchange of the electronic states repre- 
senting the normal, or stand-by, condition and the operate 
condition. For example, the single-tube gate of Fig. 4-la can 
be changed to a buffer merely by biasing both grids positively 
and furnishing input signals as negative pulses (the reverse of 
the gate conditions). Then, if either input is signaled, the plate 
current will be terminated and a positive output signal will 
result. This same transformation from gate to buffer can be 
effected on any of the other circuits, Figs. 4-16 and c and 4-26. 
The only engineering difference between a gate and a buffer is 
in the amount of energy dissipated during the stand-by periods. 

At first glance it may appear that the same results can be 
achieved by connecting both A and B to C so that either A or B 
can send a signal through to the output C. Unfortunately 
such an arrangement also connects A and B together in a " sneak 
circuit" having definitely undesirable results. For example, 
pulses originating at A and destined for C via a simple Y con- 
nection also flow into B, where they are either dissipated entirely 
or injuriously affect the apparatus connected to input B. The 
buffer connection eliminates this possibility. 

Buffers are commonly used in the input circuits of an elec- 
tronic register, e.g., in the decade rings of a device like the 
ENIAC. Here it is necessary to be able to step the ring from 
any of three inputs, a normal input, a carry-over from the decade 
of the next lower order, or a read-out input consisting of a series 
of 10 cycling pulses used to cycle the ring once for read-out. A 
three-input logical or circuit is used to accomplish this, although 
it is difficult to point to precisely the three tubes which constitute 
the buffer since this function and certain gating operations are 



SWITCHES AND GATES 



45 



combined in some of the tubes. The combination of functions 
is common when pulses traverse gates and buffers alternately, 
because a buffer may be inherent in the circuit so that additional 
tubes are not required for circuit isolation. An example of such 



OUTPUT IMPULSES 




Fig. 4-4. A combination of and and or circuits. 

a circuit is Fig. 4-4, in which the output circuit will respond if 
inputs Ai and B\ are energized or if A 2 and B 2 are energized (or, 
also, if both combinations are energized, "or" being used to 
include "and" as stated above). 

4-5. Assemblies of Gates, Buffers, and EPDCC's 

Combinations of gates and buffers with the EPDCC's described 
in Chap. 3 make possible a complete device capable of carrying 
out an elementary arithmetic operation (addition or subtraction). 
In the descriptions which follow, three examples are given in 
order of increasing complexity. Each involves high-speed 
electronic techniques which permit a complete addition in less 
than a millisecond, and each performs decimal arithmetic. These 
simple examples illustrate principles which are used in large-scale 
computing machines. 

4-5-1. A Single -digit Adder. In the upper part of Figure 
4-5a is a gate-flip-flop combination. The flip-flop opens the gate 
when a stimulus is applied on the set lead into the flip-flop. A 
stimulus on the reset lead closes the gate. The master oscillator 
generates a continuous stream of pulses at some convenient fre- 
quency such as 100 kilocycles. The decimal pulsing counter 
is a modulo 10 ring (of the 20-triode type, for example). 

A start signal causes the flip-flop to open the gate. The stream 
of pulses from the master oscillator is permitted by the open 



46 THE BASIC ELEMENTS OF MACHINE COMPUTATION 



gate to reach the pulsing counter and also to follow the lines 
shown in the diagram to other parts of the circuit. The decimal 
pulsing counter counts exactly 10 pulses, going through one 
complete cycle in doing so. When it reaches zero again, the 
tenth stage, it generates an impulse on the reset input to the flip- 
flop. This closes the gate, suppressing the passage of the 11th, 
12th, and succeeding pulses from the master oscillator. In 



master 
oscillator 



GATE 



DECIMAL 

PULSING 

COUNTER 



FLIPfFLOP 



II |i 



SET 



RESET 



START 



A 



SET 



RESET 



i r 



flip4flop 



DECIMAL 
ADDEND 
REGISTER 



DECIMAL 
AUGEND a SUM 
ACCUMULATOR 



r^ 



GATE 
(MINUS) 



GATE 
(PLUS) 



_ J 



Fig. 4-5a. An idealized single-digit decimal adder. 

short, the upper part of Fig. 4-5a illustrates a device for 
introducing exactly 10 pulses on a pulse bus each time a start 
signal is received. 

Let us consider the use of this circuit in solving a simple 
single-digit decimal addition problem involving no carry-over. 
Suppose it is required to add the single-digit decimal numbers 3 
and 5. First, the addend, 3, is stored in the decimal addend 
register, which is a ring counter of the type described in Chap. 3 
which can be set to represent a particular digit. The number 3 
is stored in this counter simply by setting the counter on its 
third position. 

The accumulator shown in the lower left-hand corner of Fig. 
4-5a is a similar ring counter, in which the augend 5 is stored. 

The start signal, as described above, permits 10 pulses from 
the master oscillator to be introduced via the bus to the addend 



SWITCHES AND GATES 47 

register and to the gate labeled gate (plus). This gate is closed, 
however, at the beginning of the operation. The pulses therefore 
do not enter. 

The addend register, as indicated above, was set on the third 
step of its cycle prior to the introduction of the train of 10 pulses. 
The first seven pulses in the train advance it through the rest of 
its cycle to zero. When the register transfers from the ninth 
step to zero, it opens the gate labeled gate (plus). With this gate 
open, the last three pulses of the train of 10 reach the decimal 
accumulator, where the number 5 has already been stored. 
These three pulses advance the accumulator from 5 to 8, which 
is the required sum. 

At the close of the operation, therefore, the sum is read from 
the accumulator where the augend had originally been set. 3, the 
number set into the addend register at the start of the operation, 
appears there again at the close of the operation, because 10 
pulses have produced a complete cycle of this register, returning 
it to 3. 

For subtraction, another gate-flip-flop combination is pro- 
vided, as shown by the dotted lines of Fig. 4-5a. Suppose it is 
required to subtract 3 from 5. 3, which is the subtrahend in this 
example, is set into the same register in which it appeared in the 
preceding example as the addend. 5, the minuend, is set into the 
register used for the augend in the addition example. The start 
pulse opens the subtract gate labeled gate (minus). It also causes 
the gate leading from the master oscillator to the counter to be 
opened. The first seven pulses reach the accumulator via this 
gate, cycling it from 5 to 2. This gate is closed after the seventh 
pulse has passed, just as the gate (plus) was opened in the preceding 
example. The required difference, 2, is thus obtained in the 
accumulator by a process which amounts to the addition to 
5 of 7, the 10's complement of 3. (The carry-over is disregarded.) 

In practice it is very difficult to operate a counter in the fashion 
described. Close timing is required when the register counter 
switches from 9 to because the same signal which effects this 
advance in the register must also be routed through the proper 
add or subtract gate. For this reason the elementary con- 
figuration shown is not actually used. Instead, a separate 
source of pulses is used to drive the add or subtract gate, and this 
stream is delayed slightly in relation to the series of pulses which 



48 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

drives the register. These two streams of pulses are accurately 
synchronized, and the delay is held constant within close limits. 
If this system is expanded to handle more than one decade, 
it is necessary to use complements with respect to 9 instead of 10 
in all but the last decade in order to carry out subtraction. How- 
ever, the register counter requires 10 pulses in order to effect a 
complete cycle. For this reason the register in a practical 



I0 S PULSES 












DECIMAL 

ADDEND 

REGISTER 




m 


_ 1 





START 



SET 1 
1 ■ T 



FLIP-FLOP 



'S PULSES 



GATE 
(MINUS) 



SUBTRACT 



GATE 
(MINUS) 



GATE 
(PLUS) 



GATE 
(PLUS) 




ADD 



DECIMAL 
ACCUMULATOR 



Fig. 4-56. A practical decimal adder element. 

machine is driven by a series of 10 pulses while the gates leading 
into the accumulator are driven by series of 9 pulses which are 
synchronized with the spaces between the 10's pulses. No 
matter how complex the computer, however, only one standard 
pulse generator is needed to drive everything else. With a 
source of accurate 10's pulses and synchronized interspersed 9's 
pulses in the computer, an elementary single-digit adder resembles 
the device shown in Fig. 4-56. In this figure the 10's pulses are 
routed to the register as before and cause this device to cycle 
once around. The 9's pulses meanwhile are reaching the two 
gates associated with the flip-flop. One of these gates is closed 



SWITCHES AND GATES 49 

from the time the start signal is received until the register transfers 
from 9 to 0, whereupon it opens. The other gate, which is used 
for subtraction, does the opposite. Two additional gates are 
shown, one each for addition and subtraction; one of these two 
is kept open throughout the operation and causes the operation 
to be either addition or subtraction. The outputs of these two 
gates are buffered together and brought to the input of the 
accumulator. The accumulator thus receives either the contents 
of the register (add) or the 9's complement of the register contents 
(subtract). This arrangement of pairs completely avoids the 
difficulty of having a set and a reset pulse reaching the flip-flop 
at almost the same time when the register flips from 9 to 0. 
This is essentially the scheme of operation of the register- 
accumulator connections for a single digit in the ENIAC, which 
will be described more completely in the third example. The 
example just given shows the use of gates for data switching 
(i.e., switching streams to indicate add or subtract) and for data 
conversion (the pulsing of the register by the 10's pulses to con- 
vert the register contents to pulses). 

4-5-2. An Elementary Parallel Adder. The second example 
illustrates a method for handling multiple-digit numbers, which 
involves carry-overs. This method is not the most efficient 
but will serve for the purpose of illustration. The combination 
of units shown in Fig. 4-6 can add two three-digit decimal 
numbers and perform the required carries. It can also take the 
difference by means of 9's complement addition. In this example 
let us assume that 379 is stored in the register and 468 is stored 
in the accumulator. They can be stored by the use of the settable 
counter or by first setting the counters to 000 and pulsing each 
digit ring the appropriate number of times. For addition, three 
distinct operations are involved. First, the units ring of the 
register is cycled with the 10's pulses as described in the previous 
example, and the output is used to control the add gate routing 
the selected pulses from the stream of 9's pulses into the units 
ring of the accumulator, thereby setting this ring to its new 
value (7). Meanwhile, another flip-flop connected to the units 
ring of an accumulator takes cognizance of the fact that this 
ring has passed from 9 to in the process of bringing up the 7, 
and it serves to open a special carry gate. After the stream of 
10 pulses has been completed, a special 11th pulse is sent to the 



50 THE BASIC ELEMENTS OF MACHINE COMPUTATION 



carry gate. Since this gate is open because of the carry-over, 
the 11th pulse can enter the 10's ring of the accumulator. The 
1 1th pulse is supplied by the same master pulse-generating equip- 
ment which supplies the 10's pulses and the 9's pulses, and it is 
properly phased with them. So far the digit 9 has been added 





















IOs PULSES 








— *- 












I0 2 


-^ 




10' 




10° 




REGISTER 




REGISTER 




REGISTER 



FLIPtFLOP 



'S PULSES 



1 



START 



FLIPtFLOP 



ADD 
GATE 



10 s 



START 



FLIP-FLOP 



ADD 
GATE 



ACCUMULATOR 



O 



CARRY 
FLIPtFLOP 



RESET 



CARRY 
GATE 



CARRY 



PULSE 



10' 



START 



ACCUMULATOR 



O 



CARRY 
FLIPtFLOP 



RESET 



CARRY 
GATE 



I0 l 



ACCUMULATOR -*J 



CARRY 
FLIPtFLOP 



RESET 



CARRY 
GATE 



Fig. 4-6. A multiple-digit decimal adder. 



to the numbers 468 in the accumulator and has produced the 
new number 477, which is correct, including the handling of the 
carry-over. The entire process is repeated except that the next 
stream of 10's pulses is routed into the 10's ring of the register, 
where they serve to rotate the 10's ring of the accumulator by an 
appropriate amount, and the 11th pulse again causes a carry-over, 
this time into the 100's ring of the accumulator. One more 
repetition of the process on the 100' s ring of the register com- 
pletes the entire operation, and the number 847 stands in the 



SWITCHES AND GATES 51 

accumulator. Switching of the input streams of 10's and 9's 
pulses would be accomplished by additional gates. 

On a desk machine, subtraction is carried out by running the 
machine wheels and carry backward. For a machine involving 
electronic counter rings this is not convenient, and some scheme 
which has the same effect must be sought. Except for carry, 
it would suffice to turn each ring forward a number of steps equal 
to 10 minus the digit appearing in the subtrahend. When carry 
is taken into account, it turns out (as every desk-calculator 
operator knows) that the lowest order digit which differs from 
is complemented with respect to 10, the higher order digits are 
complemented with respect to 9, and the zeros at the low end of 
the number, if any, are left unchanged; the resulting number is 
added to the minuend, and the final carry to the left is ignored. 
(The term complemented with respect to 10 means that the digit 
is replaced by the digit to which it must be added to give the 
number 10.) Thus to subtract 3,790 from 4,681, the comple- 
ment of 3,790 is written 6,210, and this is added to 4,681, ignoring 
the final carry, to give 0,891. Since the position of the digit to 
be complemented with respect to 10 depends on the number being 
complemented, this method is not easily applicable to automatic 
machinery. An equivalent scheme results from noticing that 
the complement obtained in this fashion exceeds by 1 the number 
which would be obtained by taking the 9's complements of all 
digits. Subtraction on many machines including the ENIAC 
is carried out in this fashion: the 9's complement of the subtra- 
hend is added to the minuend, ignoring the final carry, and 1 is 
added to the result. Thus, for the example above, 6,209 would 
be added to 4,681, ignoring the final carry, to obtain 0,890, and 1 
is added to the result to obtain 0,891. The correcting pulse 
which is used to add 1 may initiate some carry-overs. There- 
fore, after it occurs, it is necessary to operate each of the special 
carry gates once, beginning with the one of lowest order. Alter- 
natively, the digit of the lowest order might be corrected immedi- 
ately after adding the complement of the lowest order digit in 
the register. No net gain would result from this scheme since 
the same amount of time would have to be consumed when 
pulsing the digits of high order. 

Some improvement in the scheme of addition and subtraction is 
realized when carry-over is accomplished while cycling the 



52 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

counters, thus eliminating the need for an 11th pulse to follow 
later. This is analogous to the scheme used on ordinary mechani- 
cal counters, in which the input directly turns over the number of 
wheels required. The elementary parallel adder Ss a compara- 
tively slow device since it is capable of handling only one digit 
at a time. Thus, in the example cited, an addition takes 30 
pulse times. In general, if the machine is handling numbers of m 
digits on the radix r, the total time for addition in this type of 
device is rar. This becomes an appreciable factor if m is 10 
or more. The more advanced type of parallel adder described in 
the following example is faster. 

4-5-3. An Improved Parallel Decimal Adder. The method of 
addition employed in the ENIAC requires the simultaneous 
use of two complete accumulators and furnishes an example of 
an elaborate usage of gates, buffers, and EPDCC's. Altogether, 
each accumulator decade requires the use of 45 tubes (actually 
contained in 32 envelopes) to carry out the functions of addition, 
subtraction, carry-over, accumulative carry, and clearing, plus 
a number of necessary electronic operations of no direct arith- 
metic importance. Thus, to carry out a simple addition (or 
subtraction) of two 10-digit decimal numbers, 900 tubes are 
required for digit accumulators and another 14 for plus-minus 
sign indication. This does not include the common pulse- 
generating equipment and the program controls which bring 
these accumulators into play. 

Figure 4-7, a block diagram of a single ENIAC accumulator 
decade, illustrates the functions of the various tubes. One 
decade ring is used to store each digit of the addend and one to 
store each digit of the augend and later the sum. The former 
unit is in the transmit condition and the latter in the receive 
condition. In addition to the 20 tubes of the decade ring itself, 
there are the following components listed by identification 
number: 

1. Buffer for cycling the ring when the accumulator is used to 
transmit a number (add or subtract). 

2. Gate for receiving pulses when accumulator is receiving. 

3. 4, 5, 6. Pulse standardizer and driver tubes to drive the 
ring on any operation. 

7. Gate for control of carry-initiating signal. 



SWITCHES AND GATES 



53 



8. Gate for synchronizing the timing of the carry and for 
resetting the carry flip-flops. 

9, 10, 11, 12. Flip-flop and associated gates for remembering 
when a carry is required and for implementing the carry 



CARRY PULSE TO 
NEXT DECADE 



23 




\ 






CARRY-CLEAR 




' 


GATE- (RECEIVE) 


24 


- ( 


* V 



®- 




DIGITS 



SUBTRACT 
OUTPUT 



ADD 
OUTPUT 



CARRY PULSE FROM 
PREVIOUS DECADE 




Fig. 4-7. ENIAC accumulator. 



when signaled by tube 8. Also used on transmit to produce 
the add pulses and the 9's complement pulses from the 9's 
pulses. 

13, 14. Gates used to choose sum or difference for output. 

15, 16. Amplifiers. 

17, 18, 19, 20. High-power output buffers to permit the con- 



54 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

nection of many accumulator transmitters to a single 
number-transmission bus. 

21, 23. Amplifier and gate for the control of a single carry. 

22, 24. Same for accumulative carry. 

The improvement afforded by this arrangement is the increase 
in speed realized by having all decades of an accumulator pulsed 
simultaneously rather than in the sequential manner indicated 
in the previous example. Theoretically, this addition should 
require m + r cycles (as compared to mr machine cycles for the 
example previously discussed) in order to handle the possibility 
of as many carries of an m-digit number of radix r. The EN I AC 
actually requires 2r machine cycles for any value of m up to 
the maximum possible value 20. This is caused by the provision 
of two complete independent carry circuits on each decade, one 
for ordinary single carry and one for the accumulative carry 
(in which one carry may originate another, such as the addition 
of 00 • • • 001 to 99 • • • 999). In the block diagram of 
Fig. 4-7, a single carry would flow through gate 8 and tubes 21 
and 23 to the decade of next higher order. If this should create 
another carry, it will flow through 3, 4, 5, 6 of the next decade 
to the input of the ring (flipping it from 9 to 0), and gate 7 will 
cause this same pulse to pass without further delay directly to 
22 and 24 and thence to the next higher decade. The timing of 
the 10's and 9's pulses is as described in the earlier example and 
constitutes the first half of an addition time. In the second half 
the carry-over tubes 23 and 24 are opened, and gate 8 is pulsed by 
a reset pulse. Sufficient time must be allowed to permit an 
accumulative carry of maximum length and to terminate this 
operation and initiate the next one. The total time for both 
halves is 200 microseconds. The timing of the ENIAC as a 
whole is controlled by 20 pulses generated by a cycling unit in 
which a master oscillator, a 20-stage ring, and numerous gates 
and signal shapers all combine to produce a set of standardized 
pulses which are made available throughout the entire machine. 

REFERENCES 

1. Aiken, H. H., and G. M. Hopper, The Automatic Sequence Controlled 
Calculator — I, Electrical Engineering, Vol. 65, Nos. 8 and 9, pp. 384-391 
(August-September, 1946); The Automatic Sequence Controlled Calcu- 
lator— II, ibid., No. 10, pp. 449-454 (October, 1946); The Automatic 



SWITCHES AND GATES 55 

Sequence Controlled Calculator — III, ibid., No. 11, pp. 522-528 (Novem- 
ber, 1946). 

2. Burks, A. W., Electronic Computing Circuits of the ENIAC, Proceedings 
of the Institute of Radio Engineers, Vol. 35, pp. 756-767 (1947). 

3. Cornelius, E. C, Germanium Crystal Diodes, Electronics, Vol. 19, 
No. 2, pp. 118-123 (1946). 

4. Dicke, R. H., A High Speed Coincidence Circuit, Review of Scientific 
Instruments, Vol. 18, pp. 907-914 (1947). 

5. Harvard University, Staff of the Computation Laboratory, A Manual 
of Operation for the Automatic Sequence Controlled Calculator, Annals 
of the Computation Laboratory of Harvard University, Vol. I (Harvard 
University Press, Cambridge, Mass., 1946). 

6. Kurshan, Jerome, Cathode-Follower Gate Circuit, Review of Scientific 
Instruments, Vol. 18, pp. 647-649 (1947). 

7. North, H. Q., Properties of Welded Contact Germanium Rectifiers, 
Journal of Applied Physics, Vol. 17, pp. 912-923 (1946). 

8. Rossi, Bruno, Method of Registering Multiple Simultaneous Impulses 
of Several Geiger's Counters, Nature, Vol. 125, No. 3156, p. 636 (1930). 

9. Stephens, W. E., Crystal Rectifiers, Electronics, Vol. 19, No. 7, pp. 
112-119 (1946). 

10. The Stevens Arnold Co., Inc., South Boston, Mass., Millisec Relay 
(1948). Pamphlet. 

11. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, Report on the ENIAC, Technical Report I (University of 
Pennsylvania, Philadelphia, Pa., June 1, 1946). 2 vols. 

12. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, Functional Description of the EDVAC, Research Division Report 
50-9, a report of development work under contract W-36-034-ORD-7593 
with the Ordnance Department, Department of the Army (University 
of Pennsylvania, Philadelphia, Nov. 1, 1949). 2 vols. 



CHAPTER 5 
A FUNCTIONAL APPROACH TO MACHINE DESIGN 

5-1. Introduction 

In the design of a general-purpose digital computing machine, 
it is necessary to organize elementary potential digital computing 
components, switches, and gates into a device which can receive 
input information and operate upon it in a specified manner 
according to chosen systems of arithmetic and numerical analysis. 
The operations performed are initiated by information derived 
either from the data-input system or from the results of the 
arithmetic computation within the machine. This information 
is used to select specific commands from a group of operational 
commands which control the process of computation. The 
number of operational commands available and the extent of 
facilities for storage should be large enough that the arithmetic 
computations within the machine can be performed expedi- 
tiously. At the same time a minimum of circuit complexity 
is desired. 

An example of a typical arithmetic problem is presented in 
the following paragraphs. Several approaches to its solution, 
each demonstrating a method suitable for use with a machine of 
different capabilities, are set forth. The exposition is intended to 
show how a machine which can perform only a few simple arith- 
metic operations, in accordance with correspondingly simple 
instructions, can be made to solve a typical problem by following 
a set of explicit instructions. It is intended to show, also, that 
the arithmetic techniques to be followed in solving a particular 
problem efficiently on any given machine must be selected with 
regard for the capabilities of the machine. Conversely, when a 
new machine is to be designed, a consideration of its intended 
uses, i.e., of the kinds of problems it will be asked to solve, is of 
obvious importance in deciding what arithmetic operations it 
should be designed to perform. 

56 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 57 































3 
































































































1 3 
























cm • 








































































H 
















































i •« 








































































3 


























o 


^H 


o 


00 


CM 


oo 


tP 


OS 


iO 


CD 






o 


iO 


o 


CO 


o 


CM 


CD 


iO 


CM 


CD 




3 


o 


o 


i— 1 




CM 


CO 


^ 


<N 


CM 






o 


o 


o 


o 


O 


o 


O 


o 


O 


O 




s 
























<M 
























e* 
























5ss 


_ 






















1— 1 


+ 


^ 






















ub 


e* 
H 


3 






















^ 
























3 
























« 


CJ 






















3 


So 






















H 


+ 

8 
























"^ 
























« 


























o 


en 


iO 


a 


<tf 


^M 


© 


■>* 


CM 


CS 






o 


CM 


CO 


i^~ 


CO 


o 


© 


CO 


Q 


00 




5j 


o 


O 


o 


© 


I— 1 


CM 


CO 


CD 


00 


oo 




© 


o 


o 


o 


o 


O 


o 


O 


o 


o 






o 


o 


© 


o 


o 


o 


o 


^ 


© 


o 






o 


TH 


o 


00 




CO 


OS 


00 


tjh 


r^ 






o 


lO 


o 


CO 


o 


CM 


iO 


CO 


Oi 


<M 




H 


o 


O 




^H 


(M 


CO 


TF 


CM 








o 


O 


o 


o 


O 


o 


o 


o 


O 


o 




►5 


J 


T— 1 


<M 


CO 


^ 1 


iO 


CD 


t^ 


00 


OS 


© 



58 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

6-2. An Example 

Suppose x and y denote two measurable physical quantities, 
such as the orthogonal components of a velocity which varies in 
magnitude and direction but remains at all times in the plane 
of x and y. A large number of measurements of corresponding 
values of x and y have been made. The problem is to compute 
the scalar magnitude of the velocity vector for each pair of x 
and y values, i.e., to compute the square root of the sum of the 
squares of corresponding x's and y's. Table 5-1 contains 10 such 
pairs of values of the measured quantities, in columns headed 
x and y. This table contains also the value of u, where 



u = -\/x 2 + y 2 

for each pair of x and y values. The meaning of each of the blank 
columns in the table will be made clear shortly. 

Suppose the values of u are to be obtained by means of cal- 
culations performed on a desk machine which can perform 
addition, subtraction, multiplication, and division directly but 
has no automatic device for extracting the square root of a 
number. To obtain the square root of x 2 + y 2 , therefore, some 
method of approximation will be required. (Chapter 7 dis- 
cusses such methods, and Chap. 8 contains a summary of the 
properties of various commercially available desk calculators; 
no detailed discussion of either of these topics is of primary 
importance here.) Let us also suppose two additional factors: 
(1) the value of y shown in the table for the first measurement is 
known to be exactly zero as a consequence of the way the experi- 
ment was run, and (2) the measurements were made frequently 
enough so that succeeding values of u are known to lie close 
together. Because succeeding values do lie close together, each 
computed value of u may be used as the assumed value in the 
next step of the computation. u k , the kth value of u, may be 
found by using the formula 

_ [Q 2 + y 2 )/u k -i] + Uk-i 
Uk ~ 2 

The detailed instructions for carrying out the computation 
expressed by this formula may be given to a machine operator 
in the following form: 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 59 

1. For measurement No. 1 copy the number x into the column 
headed u. 

2. For each of the other measurements, in order, multiply the 
number x by itself; add to the product, without removing 
it from the machine's register, the product of the number y 
multiplied by itself. This leaves x 2 + y 2 in the product 
register. 

3. Clear the quotient register, leaving the accumulated product 
in the machine, and divide this product by the entry in the 
u column computed for the next earlier measurement, or 
u k -i. 

4. Transfer the quotient to the product register and add to it 
the entry from the u column computed for the next earlier 
measurement, u k -i. 

5. Clear the quotient register and divide the number in the 
product register by 2; write the result to four decimal 
places in the column headed u. 

5-2-1. Variations in the Method. In following the set of 
instructions given above the operator had to fill in only the 
answer column, headed u, in Table 5-1. The use of the other 
columns is illustrated below in an example of the solution of the 
same basic problem on a different calculator. Variations in the 
arithmetic techniques and, consequently, in the program of 
instructions are required because this calculator is not capable 
of performing all the operations contained in the foregoing pro- 
gram. Analogous variations are required when the same basic 
problem is handled on two different large-scale machines. 

At the end of step 2 in the calculation described above, the 
quantity x 2 + y 2 had been built up in the product register. The 
process of building it up depended upon the ability of the machine 
to multiply and to accumulate products. It might be more 
convenient to build a machine which would not accumulate 
products. In this case, if the machine could multiply, add, and 
divide, but all in different components, it might be necessary to 
write down the quantities x 2 , y 2 , and x 2 + y 2 as they are com- 
puted so that they can be reinserted in the proper places in the 
machine at the desired time. For such a machine, the procedure 
would be modified in the manner stated before the directions 
were turned over to the operator. Furthermore, the x 2 column 
of Table 5-1 would be filled during the computation. 



60 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

A more serious modification is required if the machine will 
not carry out some of the arithmetical operations used. Suppose, 
for example, that the machine will not divide easily. There 
are two divisions in the process, and they would have to be 
replaced. The second is division by 2, and it is simply replaced 
by multiplication by 0.5000. The first is division by the next 
earlier value of u; this is more difficult. Methods described in 
Chap. 7 permit an easy running computation of a slowly changing 
value for vr 1 , using only the operations of multiplication and 
subtraction. It is convenient to use the fact that the first x 
is exactly 1 and the first y is exactly 0. Then the set of instruc- 
tions or commands for a machine incapable of division would have 
to be changed from the set above to a set similar to these: 

V. For measurement No. 1, write the number 1 in the column 
headed u and the column headed vr 1 . 

2'. For each of the other measurements, in order, multiply 
the number x by itself; add to the product, without remov- 
ing it from the machine's register, the product of the 
number y multiplied by itself. This leaves x 2 + y 2 in the 
product register. 

3'. Transfer the number in the product register to the key- 
board, clear the product register, and multiply the number 
by the entry vr 1 for the next earlier measurement, vj^} v 

4'. Add the entry u for the next earlier measurement, v k -i, to 
the number in the product register. 

5'. Transfer the number in the product register to the key- 
board, clear the product register, and multiply by 0.5. 
Write the result to four decimal places in the column 
headed u. 

6'. Clear the machine and insert the number 2.00000000 in the 
product register and subtract from it the product of u 
(just calculated) and vr} v 

7'. Transfer the number in the product register to a keyboard 
(the multiplier keyboard if the machine is equipped for 
automatic multiplication), clear the product register, and 
compute the product of the number just transferred and 
the number uj~} v Enter the product to four decimal 
places under vr 1 . 

This method of performing division was first brought to our 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 61 

attention through its use in the Mark II Calculator. 3 It may 
be stated thus: 

u~ k l « (2 - u • u^uj^i 

but it is not of primary importance in this discussion. 

The commands to the operator in this set still make use of 
some of the properties which are built into the machine and which 
might not be convenient in a machine built on the fast operating 
principles being described here. It might be true, for example, 
that a convenient machine would not immediately operate on 
the result of any other operation; in this case, the properties of 
the product register used repeatedly in the list of commands 
could not be used to permit double operations like 6', and the 
single operation 4' would become more complex. Although no 
machine yet proposed has been reduced to the point that the 
result of an elementary arithmetic operation is immediately 
removed from the computing component, this is certainly a 
possibility which must not be ignored. If this were the case for 
a machine with a human operator and if the machine were capa- 
ble of addition, multiplication, and subtraction but not of 
division, then a set of fairly stereotyped commands would be 
written, and all the spaces in Table 5-1 except a few for measure- 
ment No. 1 would be filled. The following commands or their 
equivalent would be required: 

1" . Enter the number 1 in the u column for measurement 

No. 1. 
2". Enter the number 1 in the u~ 1 column for measurement 

No. 1. 

For each other measurement in succession proceed as follows: 

3". For entry in x 2 column, multiply; multiplicand is entry 
in x column; multiplier is entry in x column. 

4". For entry in y 2 column, multiply; multiplicand is entry 
in y column; multiplier is entry in y column. 

5". For entry in x 2 + y 2 column, add; augend is entry in x 2 
column; addend is entry in y 2 column. 

6". For entry in (x 2 + y 2 )/uk~i column, multiply; multipli- 
cand is entry in x 2 + y 2 column; multiplier is entry in 
ir 1 column for next earlier measurement, u^} v 



62 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

7". For entry in 2u column, add; augend is entry in (x 2 + y 2 )/ 

Uk-\ column; addend is entry in u column for next earlier 

measurement, u k -\. 
8". For entry in u column, multiply; multiplicand is entry 

in 2u column; multiplier is the number 0.5000. 
9". For entry in u • %l\ column, multiply; multiplicand is 

entry in u column; multiplier is entry in u~ x column for 

next earlier measurement, u^ v 
10". For entry in 2 — u • u^li column, subtract; minuend 

is the number 2.00000000; subtrahend is entry in u • uj^i 

column. 
II". For entry in u~ l column, multiply; multiplicand is entry 

in 2 — u ■ u^li column; multiplier is entry in u~ l column 

for the next earlier measurement, uj^} v 

The stereotyped nature of the commands renders an advantage 
in clarity and in ease of writing a program. It should be noted 
that each command except for the unnumbered command "For 
each other measurement in succession:" consists of a prescribed 
arithmetical operation, addition, subtraction, or multiplication, 
in which the two entering numbers are specified by their position 
in Table 5-1 and in which the disposition of the resulting number 
into a position in Table 5-1 is specified. Burks, Goldstine, and 
von Neumann 2 have pointed out the advantages of establishing 
a simple stylized form of this general type no matter what the 
characteristics of the machine. For the convenient operation 
of a general-purpose machine, they point out, it is essential that 
some steps be taken to translate the nonconforming command 
quoted above to the same stereotype form. This translation of 
description of all possible operations to prescribed forms has been 
called the logic of the machine by these authors, and the term 
is now in general use. 

5-3. Machine Requirements 

The earlier part of this chapter describes a simple calculation 
in which various postulated types of machinery operated by a 
human operator are used. The two preceding chapters describe 
techniques which seem to promise arithmetic calculation with 
the elementary operations requiring only microseconds for com- 
pletion. For a machine using these new components, a human 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 63 

operator intervening at every command would encumber the 
machine with a slow-moving element whose reaction time is 
intolerably long. For example, during a second of time the 
machine is capable of carrying out 1,000 multiplications, but the 
operator is capable of issuing at most a single command. It is 
essential that a sequencing mechanism similar to, but faster 
than, the human operator be built into the machine if its potential 
speed is to be utilized. 

An automatic sequence-controlled calculator is a computing 
machine into which such a mechanism is built. It follows a pre- 
scribed set of explicit instructions automatically. 

Such a machine is usually much faster and more accurate than 
a twelve-year-old arithmetic student, but less versatile. Its 
lack of versatility can be offset by its speed because combinations 
of a few simple operations can be made to effect results equivalent 
to those which would be produced by the twelve-year-old in 
response to less explicit instructions. These combinations of 
simple operations are therefore components of computation. 

A sheet of paper, Table 5-1, was also included as a component 
used in the calculation described above. This sheet of paper 
is so inexpensive that it is overlooked frequently as a component, 
but those experienced in calculation know well that most of the 
time required in calculation and essentially all the errors of the 
calculation are connected with the reading of data from and the 
entering of data on this sheet of paper. The potentialities of 
speed of computation with the new components and the resulting 
extent of calculations which can be made in a reasonably short 
time demand a storage medium which is faster and more accurate 
than pieces of paper marked and read by a human operator. 
It is also obvious that when complicated calculations are per- 
formed in small increments, a voluminous storage outside the 
arithmetic units is required to store the partial results until they 
are pieced together into a final solution. The voluminous 
storage to which quick access must always be available is at 
present a chief consideration in connection with computing 
instruments. Many machines have several different storage 
units, some of which are more quickly accessible than others. 

Finally, implicitly, the problem examined above had input data 
furnished by some source and output data read by some inter- 
preter. It is true that these input data appeared on the basic 



64 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

storage medium of the machine (Table 5-1) and that they were 
read from this basic medium by the final interpreter, but this is 
not necessarily a general state of affairs. It is likely that a 
sufficiently fast and accurate storage component for a machine 
of the speed considered here will be an integral part of the 
machine, not a detachable part. Thus some convenient attach- 
able medium for input of basic data must be provided, and some 
convenient detachable medium for output of results must also 
be provided. 

5-4. Stereotyped Commands for the General -purpose Machine 

For the sample problem, somewhat similar to some problems in 
data reduction, a set of 11 commands of a stereotyped nature, 
designated by 1" to 11", and one command not stereotyped 
located between commands 2" and 3" were written above. If 
these commands are carried out in sequence, the data will be 
reduced as required. These commands have the advantages 
that a few commands govern a large number of calculation steps 
and that the commands themselves are all similar (except for 
the one nonconforming command); specifically, their similarity 
is that each specifies an arithmetic operation, and each specifies 
where the numbers entering into this operation can be found and 
where to dispose of the number generated by this operation. 
These commands have disadvantages in that they were written 
down with no particular regard for the engineering aspects of 
the computing components and they include the nonconforming 
command. 

A set of 43 commands is listed below. Each is of a stereotyped 
nature, is written with some regard to ease of engineering achieve- 
ment, and is so chosen that the sequence of 43 commands controls 
the same set of calculations as the earlier sequence. In this 
sequence there are no nonconforming operations, and every 
operation has a stereotyped form. This form specifies a desig- 
nator for the operation, a description of the operation, and the 
location of one number involved in the operation. No operation 
requires further specification. In an engineering realization, 
the designator does not turn out to be a part of the command 
itself but is a statement of where the command is to be placed 
in the machine. It is assumed that the location of the number 
involved in the operation is itself a numerical code with 10 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 65 

assigned values. Thus, every command can be written as a 
number, a few digits of which state the code number describing 
the prescribed operation and the remaining significant digits of 
which state the address code number of a number involved in the 
operation. In addition, the designator, or the place in the 
machine at which this command will be placed, must also be 
specified. 

A machine for which commands can be written in this way is 
clearly a versatile machine. Since the commands are numbers 
and since the machine must expect to have numbers introduced 
to it in some fashion, it certainly should not be difficult to enter 
these coded commands into the machine. Since a small number 
of commands will control a lengthy calculation, the amount of 
time spent in writing these commands down for insertion into 
the machine is not intolerably long. Since only a few different 
operations are required, the engineering structure of the machine 
may be expected to be attainable fairly simply. Commands 
of this sort have been described most completely and lucidly by 
Burks, Goldstine, and von Neumann. 2 They describe a set of 
21 operations which suffice for conveniently carrying out most 
problems from a large field. The authors describe the machine 
on which these commands will be carried out specifically enough 
to give a general idea of the engineering considerations involved. 

Bloch, Campbell, and Ellis 1 have described a set of commands 
more nearly of the type used in the list 1" to 11". This publica- 
tion may be more readily available than the Burks, Goldstine, 
and von Neumann paper, but the description is not as complete 
in some respects. On the other hand, a special checking scheme 
and some other novel ideas are discussed by Bloch, Campbell, 
and Ellis. 

The commands which will be used below are a set taken essen- 
tially from the set proposed by Burks, Goldstine, and von Neu- 
mann. 

5-4-1. Description of the Machine. The input and output of 
the machine will be ignored in this description. The machine 
will possess a storage element, an arithmetic element, and a 
control element. The arithmetic element will contain two main 
units: the accumulator, which will be denoted by the letter A, 
and the multiplier register, which will be denoted by the letter R. 
Addition can take place in the accumulator. It will be assumed 



66 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

that these elements can handle numbers of the sizes which will 
occur (although the accomplishment of this assumption may 
not be perfectly straightforward). The positions in which 
numbers can be stored will be addressed by numbers; thus, an 
address is an ordinal number of a position in storage, at which 
position a number involved in the calculation is to be found or 
to be placed. It will be assumed that the storage facilities are 
adequate for the problem in hand and that the size of the arith- 
metic unit is sufficiently great to carry out the calculations which 
will later be seen to be involved in modifying the commands. 
(Specifically the arithmetic unit must handle numbers as large 
as the largest ordinal number occurring in connection with the 
storage.) The commands which the machine is capable of 
following are described below in the order of their appearance in 
Table 5-2, starting on page 69. 

The first operation that the machine will be required to carry 
out is the clearing of the accumulator and the insertion in the 
accumulator of the number at a designated address. This 
operation occurs first in command V" below. The machine 
will be built so that the number placed into the accumulator 
remains unchanged in its storage position. The operation will 
not affect R. 

The second operation is the inverse of that above; it is the 
operation of storing at a specified address the number which is 
in the accumulator. This operation destroys any number 
previously stored at that address, leaves intact the number in A, 
and does not affect R. This operation occurs first in command 
2'" below. 

The third operation is the operation of destroying any number 
which is in R and replacing it by the number at a specified 
address; this operation occurs first in command 3'". The 
number remains unchanged in its storage position, and the 
operation does not affect A. 

The fourth operation causes the number in the accumulator to 
be destroyed and the product of the number in the R register, as 
multiplier, and the number at a specified address, as multiplicand, 
to be inserted in A. This operation occurs first in command 
4'". The number at the specified address remains unchanged; 
the number in R will be destroyed during the operation. 

The fifth operation retains the number in A and adds to it the 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 67 

number at a specified address. It occurs first in command 8'". 
The addend remains unchanged in its storage position. The 
operation does not affect R. 

The sixth operation is a subtraction operation similar to the 
one above. The number in A is retained, and the number at a 
specified address is subtracted from it, the difference being kept 
in A. The operation occurs first in command 21'". The 
number at the specified address remains unchanged at that 
address, and the operation does not affect R. 

The seventh operation occurs first in command 28'" below. 
To understand it, it is necessary first to have described the normal 
sequencing of commands. A program like the one written 
below is set up in a fixed sequence. The machine automatically 
carries out each command in the order of this sequence unless a 
command tells it to depart from this order. (Bloch, Campbell, 
and Ellis 1 describe a different procedure for specifying the order 
in which commands are to be carried out.) Thus, a few com- 
mands must be used to change the order in which commands 
are carried out. If these commands were not included, it would 
be impossible to recycle, and it would be necessary to write one 
command for every arithmetical operation the machine is to 
carry out. This seventh command is one used for altering the 
sequence conditionally; if the number in the accumulator is a 
negative number, there is no alteration in sequence, and the 
machine next carries out the next command (29'"), but if the 
number in A is nonnegative, the machine passes to the command 
specified in the address portion of command 28"'. 28'" is called 
a conditional transfer command or conditional jump. 

The description of this operation and of one or two others 
below is made more precise by specifying that the commands 
themselves will be placed in the storage unit of the machine. 
This is possible because the commands can be written as numbers. 
It was pointed out above that all that is required is that the 
operations to be carried out be assigned code numbers and that a 
complete command be written as a code number specifying an 
operation followed by one specifying a numerical address. The 
digits of these two parts may be run together to form a single 
multidigit number. The description is complete when it is 
stated that the designator of the command is simply the address 
in the storage system at which the command is stored. Sue- 



68 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

cessive commands in the sequence of commands will be stored 
at successively numbered addresses. Thus, the control unit 
will be required to start at a specified place and to withdraw 
commands from storage. After each command is carried out, 
the control unit passes to the next higher numbered storage cell 
for the next command unless it has been instructed by the com- 
mand just carried out to depart from the standard command 
sequence. Thus recycling and alternate sequences are made 
possible. 

The eighth operation is similar to the second except that the 
information transferred from A to the storage unit is restricted to 
be information concerning the address of a command. No 
matter what is in the accumulator, the portion of the stored 
command having to do with the operation to be carried out will 
not be changed; the only change is a change in address. This 
is an operation which is of fundamental importance in removing 
the nonconforming command of the set described earlier. The 
operation first occurs in command 31'". 

The ninth operation of the set is one which instructs the control 
to depart from the standard sequence of commands. The nature 
of its utility is apparent from its use in command 43'" below. 
Nothing in the arithmetic unit or the storage unit is affected by 
this command, which is called an unconditional transfer. 

The final operation of the set appears in storage at cell j. 
It tells the machine to stop calculations. Without this command 
the machine at the end of a calculation would proceed with 
meaningless and possibly harmful calculations. 

5-4-2. The Coding. A coding which uses the operations 
described above can be written immediately from the set 1" 
through 11". All numbers involved will be assumed to have 
been inserted in storage. Addresses in storage will correspond 
to specification of position (measurement number and column 
heading) on the work sheet referred to earlier. A block of 
storage addresses must be assigned to each column in which are 
written data which must be retained permanently; these are 
the columns headed x, y, and u. The other numbers (which 
were never written down at all in the program 1 through 5 above) 
are assigned space as necessary during the time they must be 
saved; these numbers are then discarded in favor of their suc- 
cessors. Finally, a block of consecutively numbered addresses 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 69 
Table 5-2. Commands 



Designator 


Operation 


Address 


Remarks 


V" 


Clear A and insert the num- 








ber at 


c 




2"' 


Store number now in A at 


d 


Completes 2" 


3'" 


Clear R and insert the number 








at 


a 




4/" 


Clear A and insert in it the 
number got by multiplying 
by the number in R the nurriT 








ber at 


a 




5"' 


Store number now in A at 


e 


Completes 3" 


6'" 


Clear R and insert the number 








at 


b 




fjrrr 


Clear A and insert in it the 
number got by multiplying 
by the number in R the num- 




Completes 4" 




ber at 


b 




8"' 


Add to the number in A the 








number at 


e 




9'" 


Store number now in A at 


e 


Completes 5" 


10'" 


Clear R and insert the number 








at 


d 




11'" 


Clear A and insert in it the 
number got by multiplying 
by the number in R the num- 




Completes 6" 




ber at 


e 




12"' 


Add to the number in A the 








number at 


c 




13"' 


Store the number now in A at 


e 


Completes 7" 


14"' 


Clear R and insert the number 








at 


f 




15"' 


Clear A and insert in it the 
number got by multiplying 
by the number in R the num- 








ber at 


e 




16"' 


Store the number now in A at 


c + 1 


Completes 8" 


17"' 


Clear R and insert in it the 








number at 


d 




18"' 


Clear A and insert in it the 
number got by multiplying 
by the number in R the num- 








ber at 


c + 1 




19'" 


Store the number now in A 




Completes 9" 




at 


e 





70 THE BASIC ELEMENTS OF MACHINE COMPUTATION 
Table 5-2. Commands. (Continued) 



Designator 


Operation 


Address 


Remarks 


20'" 


Clear A and insert in it the 








number at 


.9 




21'" 


Subtract from the number in 








A the number at 


e 




22'" 


Store the number now in A 




Completes 10" 




at 


e 




23'" 


Clear R and insert in it the 








number at 


d 




24'" 


Clear A and insert in it the 
number got by multiplying 
by the number in R the num- 








ber at 


e 




25'" 


Store the number now at A 




Completes 11" 




at 


d 




26"' 


Clear A and insert the num- 








ber at 


h 




27"' 


Subtract from A the number 








at 


i 




28'" 


If the number in the accumu- 




Checks to see whether 




lator is negative, pass to the 




calculation is com- 




next command in this se- 




plete 




quence; if it is not negative, 








jump to the command stored 








at 


J 




29"' 


Clear A and insert the number 








at 


h 




30'" 


Add to the number in A the 








number at 


c 




31'" 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


3"' 




32'" 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


4 ,„ 




33'" 


Store the number now in A at 


h 




34'" 


Add to the number in A the 








number at 


k 




35"' 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


6"' 





A FUNCTIONAL APPROACH TO MACHINE DESIGN 71 
Table 5-2. Commands. (Continued) 



Designator 


Operation 


Address 


Remarks 


36"' 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


7'" 




37'" 


Add to the number in A the 








number at 


I 




38"' 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


12'" 




39'" 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


30'" 




40'" 


Add to the number in A the 








number at 


c 




41'" 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


16"' 




42'" 


Store the number in A as the 
address part (leaving the 
coded operation unchanged) 








of the command 


18'" 




43'" 


For the next operation pro- 




Repeat command 3'" 




ceed in order after jumping 




and all following. 




to command 


3'" 


This command re- 
places the unstereo- 
typed command be- 
tween 2" and 3" on 
the earlier list. 



is assigned to the 43 commands V" through 43'". The control 
unit causes the commands to be carried out in sequence except 
where deviations are commanded by one of the sequence of 
commands. 

Specifically, in summary, the storage unit is loaded as follows: 
The numbers x are stored at consecutively numbered addresses, 
with the first at address a. The numbers y are stored at con- 
secutively numbered addresses with the first at address b. The 
numbers Uk as computed will be placed in consecutively numbered 



72 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

addresses with the first at address c. Suppose that there are n 
measurements; then the last x is stored at a + n — 1, etc. A 
block of 43 consecutive addresses is assigned to the coded com- 
mands V" through 43'". Special numbers are stored at address 
d and eight other addresses designated by letters e through I 
according to the scheme below : 

Address Number 

c 1 

d Latest entry for u~ l 

e Latest entry for x 2 , x 2 + y 2 , 2u, u • wjT.li, and 2 — u • Wj^}_ x in order 

/ 0.5 

g 2 

h The number a, to be replaced by the current address 

i The number a + n 

j The coded value of the command stop calculation 

k The number b — a 

I The number c — b 

Note that command I" of the early set has been carried out 
already through these storage operations. 

5-5. Plan for General-purpose Machine 

The example above indicates a plan for the convenient opera- 
tion of a general-purpose machine. The operational requirement 
is simply that a few written commands can conveniently control 
a lengthy calculation. The realization of the requirement 
suggested is by means of a few stereotyped command forms, 
each of which specifies an operation and one or more storage 
positions. The arithmetic element must be able to carry out 
the operations contained in these stereotyped commands. The 
control element must go through the pertinent commands in 
order, translate the coded description of the operations to pulses 
along busses which will cause the arithmetic elements involved 
to carry out the specified operations, and direct numbers to and 
from the specified storage locations. 

The requirement that a few written commands direct a long 
calculation is attained by (1) the inclusion of facilities permitting 
the reuse of commands with the operations unchanged and the 
addresses changed if desired; and (2) an operation permitting a 
departure from the normal sequence to a secondary one (con- 



A FUNCTIONAL APPROACH TO MACHINE DESIGN 73 

sisting of a single conditional transfer command in the example), 
the departure to depend on the calculation. 

REFERENCES 

1. Bloch, R. M., R. V. D. Campbell, and M. Ellis, Logical Design of the 
Raytheon Computer, Mathematical Tables and Other Aids to Computation, 
Vol. 3, No. 24, pp. 286-295, 317-322 (October, 1948). 

2. Burks, Arthur H., Herman H. Goldstine, and John von Neumann, 
Report on the Mathematical and Logical Aspects of an Electronic Computing 
Instrument, 2d ed., Part I: Preliminary Discussion of the Logical Design 
of an Electronic Computing Instrument (Institute for Advanced Study, 
Princeton, N.J., 1947). 

3. Harvard University, Staff of the Computation Laboratory, A Manual of 
Operation for the Automatic Sequence Controlled Calculator, Annals of the 
Computation Laboratory of Harvard University, Vol. I (Harvard Univer- 
sity Press, Cambridge, Mass., 1946). 



CHAPTER 6 
ARITHMETIC SYSTEMS 

6-1. Introduction 

Most modern arithmetics are founded upon a system for 
representing the positive integers in terms of systematic symbols 
involving the concept of carry-over. That is, they depend upon 
a systematic representation of the counting process. These 
arithmetics are the basis of all automatic computing operations 
now used. 

The basic concepts used in the more usual arithmetics have 
been described in Carmichael. 3 A detailed exposition will be 
given here of the concepts of number theory which may be 
applied directly to an understanding of computer arithmetic 
systems. 

6-2. The Fundamental Counting Systems 

The recording of counted numbers in the systems to be studied 
here will be by means of an infinite ordered sequence of marks, 
of which the first will be considered to be at the right-hand end 
and the others to proceed in order to the left. The nature of the 
sequence will be defined inductively. In each position the 
mark (or a mark with an equivalent meaning) will be admissible 
as a mark, and usually one or more other marks will be admissible. 
For example, in the decimal system the marks admissible in 
each position are the marks 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9. The 
admissible marks will not depend upon the marks present in 
other positions. For each position, the admissible marks have 
an order assigned, and the mark precedes all the rest. (The 
order of the decimal marks is that in which they are given above.) 

The rule for denoting a number is given in the following parts: 

1. The number zero is denoted by a sequence each of whose 
elements is the mark 0. 

2. To pass from the representation of any number to that 

74 



ARITHMETIC SYSTEMS 



75 



of the next higher, the first mark which is not the last 
admissible mark for its position is replaced by the next 
admissible mark in the set established for its position and 
all lower order marks are replaced by O's. 

6-3. Examples of Counting Systems 

The decimal system of counting follows this rule, with the 
understanding that a final string of O's in the sequence, extending 
to the left, is not written. As noted above, the ordered set of 
marks which may be used in each position is the set 0, 1, 2, 3, 4, 
5, 6, 7, 8, 9. The number 57 is, according to the rule, followed 
by 58, for 7 in the first (right-hand) position is the first mark 
which is not the last admissible one in its position and it is 
replaced by 8, the next higher mark. The number 26,999 is 
followed by 27,000, for the first three marks are all the last 
admissible marks in their positions, so the fourth one (which is 
not maximal) is increased by one and the last three are replaced 
by O's. The binary number system admits the marks and 1 at 
each position, and no others. Thus the counting rule gives the 
equivalences shown in Table 6-1 between the first 36 decimal and 
binary numbers. In each case, O's after the last nonzero mark 
are omitted except for the number zero. 

Table 6-1 



Decimal 


Binary 


Decimal 


Binary 


Decimal 


Binary 








12 


1,100 


24 


11,000 


1 


1 


13 


1,101 


25 


11,001 


2 


10 


14 


1,110 


26 


11,010 


3 


11 


15 


1,111 


27 


11,011 


4 


100 


16 


10,000 


28 


11,100 


5 


101 


17 


10,001 


29 


11,101 


6 


110 


18 


10,010 


30 


11,110 


7 


111 


19 


10,011 


31 


11,111 


8 


1,000 


20 


10,100 


32 


100,000 


9 


1,001 


21 


10,101 


33 


100,001 


10 


1,010 


22 


10,110 


34 


100,010 


11 


1,011 


23 


10,111 


35 


100,011 



The binary system occurs frequently in considerations of 
automatic computers, for many of the basic circuits, such as the 



76 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

flip-flop, naturally assume two stable states and are therefore 
convenient to use with binary numbers. 

In the octal system, any one of eight different marks is admis- 
sible at each position, e.g., 0, 1, 2, 3, 4, 5, 6, 7 or the binary 
equivalents of these same familiar digits, 000, 001, 010, Oil, 
100, 101, 110, 111. This system is of some interest wherever 
binary numbers occur because conversion between binary and 
octal numbers is trivially easy, and octal numbers make a fairly 
efficient set for printing and reading. These points are exhibited 
in greater detail in Sec. 6-6. 

The biquinary system was used as a convenient equivalent 
of the decimal system in some machines produced by the Bell 
Telephone Laboratories. 1 This application is attributed to 
Dr. George Stibitz. The system is one in which the decimal 
system is factored and written as a system in which the odd- 
numbered positions have five admissible marks and the even- 
numbered positions have two admissible marks. The admissible 
marks for the odd positions are 0, 1, 2, 3, and 4, and for the even 
positions and 5. The equivalence of this system with a decimal 
system should be obvious, for the sum of a pair of successive 
marks can be considered to be a decimal mark itself, there being 
10 such pairs. For example, the decimal numbers 4 and 8 
are represented in the biquinary notation as 04 and 53, respec- 
tively. Carry-over occurs at 10 as in the decimal system. 

This arithmetic system is much like the tetraquinary system 
of counting of the Mayan Indians, who had at least two systems 
of counting, one for general usage and at least one other for dates. 
These have been described by Morley 4 and by Spinden. 7 The 
Mayans used a true vigesimal system with a systematic rule of 
forming the marks. It is said that the Mayans used their 
vigesimal system prior to the time the decimal system achieved 
general usage in Europe. In this system there were 20 marks 
for each position; however, each mark (except zero) consisted 
of from zero to four dots placed over from zero to three bars. 
This system was obviously equivalent to a system in which the 
odd-numbered positions have five marks (the dot configurations) 
and the even-numbered positions four (the bar configurations). 
The Mayan zero, a " shell/' was used in the formation of 20, 40, 
etc. Although the Mayans used the system as a true vigesimal 
number system, they counted as if it were a tetraquinary system: 



ARITHMETIC SYSTEMS 



77 



the number of bars represented the number of hands and feet 
they had used up in the count, and the number of dots repre- 
sented the number of odd digits on the next hand or foot. 



Table 6-2 



Decimal 


Mayan 


Decimal 


Mayan 



1 


• 

• • 

• c • 

• • • • 


10 
11 
12 
13 
14 
15 
16 
17 
18 
19 




• 


2 
3 
4 
5 


• • 


• • • 


• • • • 





g 


• 


• 


7 


• • 


• • 


g 


• • • 


• • • 


9 


• • • • 


• • • • 









The monetary system used in the United Kingdom may be 
thought of as another, more complicated counting system, 
satisfying the rule of Sec. 6-2. The mark in the first position 
gives the number of pence, the mark in the second position gives 
the number of shillings, and the mark in the third position gives 
the number of pounds. It is common practice to separate the 
different positions by a diagonal (/), writing 26/14/11 for 26 
pounds, 14 shillings, 11 pence. There are only three positions. 
Assuming that the penny is the unit (i.e., ignoring the halfpenny), 
the marks which are admissible in the first position are the decimal 
numbers from through 11 in order; those admissible in the 
second position are the decimal numbers through 19 in order, 
and the infinite ordered set of nonnegative decimal integers is 
the admissible set for the third position. 

6-4. Three Fundamental Theorems 

The main principles of behavior of numbers under different 
number representations may be characterized by three elementary 
theorems. For all these theorems it will be assumed that there 
are ni different admissible marks for the first position, n 2 for the 
second, n z for the third, etc., with m for the ith.. The theorems 
are generally trivially true for any n% which become infinite, and 
these cases will be disregarded (e.g., the case of the British 
monetary system). 



78 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

Theorem 6-1. The number of different numbers which can be 
represented by configurations with marks different from in at 
most the first k positions is the product 

Pk = n 1 n 2 n 3 • • • n k (6-1) 

For proof of Theorem 6-1 see Sec. 6-13-1. 

From this theorem, it follows that exactly 1,000 (10 X 10 X 10) 
different numbers can be represented with three decimal digits; 
these are the numbers from through 999, inclusive. Similarly, 
to represent 1,000 different numbers with binary digits, 10 digits 
will be required, for 1,000 lies between 2 to the ninth and 2 to 
the tenth powers (512 and 1,024, respectively, expressed in 
decimal numbers). 

Theorem 6-2. In the representation of a number x, if pi 
represents the number of marks admissible which precede the mark 
actually occurring in the first position, if p 2 is the number of marks 
admissible for the second position which precede the mark in that 
position, etc., with pi the number of admissible marks for the ith 
position preceding the ith mark, then, conforming with the notation 
for P of Eg. (6-1), 

x = pi + P2P1 + P3P2 + • • • + PiPi-i (6-2) 

For proof see Section 6-13-2. Theorem 6-2 can be used to write 
the value in any known arithmetic of a number expressed in 
any other arithmetic. Thus, for example, the binary number 
10,001,001 may be expressed in decimal notation by noting that 
according to the rule it must equal 

1 • 2 7 + • 2 6 + • 2 5 + ■ 2 4 + 1 • 2 3 + • 2 2 + • 2 1 + 1 

= 128 + 8 + 1 = 137 

Similarly, in a ternary system which has the admissible marks 
0, 1, and 2, the ternary number 12,021 has an equivalent decimal 
value 

1 • 3 4 + 2 • 3 3 + • 3 2 + 2 • 3 1 + 1 = 81 + 54 + 6 + 1 = 142 

Theorem 6-3. If x is a nonnegative number and if m, pi, and 

Pi all have the significance defined above for some system of notation 
in which x is expressed, then pi is the remainder when x is divided 
by rii, p 2 is the remainder when this integral quotient (without 
remainder) is divided by n 2 , etc.: explicitly, there exists a set of 



ARITHMETIC SYSTEMS 

nonnegative integers q , q h Q2, . . . , Qi such that 

q = x = riiqi -f pi; < pi < rii 

qi = n 2 q 2 + Pt\ < p 2 < n 2 

q< = n i+ iq i+ i + Pi+i] < p i+1 < n i+1 



79 



(6-3) 



These equations uniquely determine pi. For proof see Sec. 
6-13-3. 

Theorem 6-3 furnishes a method for writing the expression for 
any positive integer in terms of any system of marks, carrying 
out the arithmetic in any known system. For example, to 
translate the number 137 to the binary system, where every 
rii = 2, repeated division by 2 suffices, as shown in Table 6-3. 

Table 6-3 



i 


m 


(ffi-i) 


(Pi-i) 


1 


2 


137 




2 


2 


68 


1 


3 


2 


34 





4 


2 


17 





5 


2 


8 


1 


6 


2 


4 





7 


2 


2 





8 


2 


1 





9 


2 





1 



Table 6-3 was calculated as follows: First all the numbers n, 
were written in their column; then the number 137 was entered 
in the center column. The division was performed qi-i/rii, and 
the quotient written directly below with the remainder written 
in the right-hand column. The binary representation is written 
by copying the right-hand column, the upper mark being placed 
at the right in copying: 

Decimal 137; binary 10,001,001 

A warning is in order at this point: this rule stated in Theorem 6-3 
cannot be applied blindly to fractional numbers expressed, for 
example, with a decimal point. 



80 THE BASIC ELEMENTS OF MACHINE COMPUTATION 



6-5. Rules of Arithmetic for Numbers with a Single Radix 

If all the numbers n,- of a system of notation like those described 
above are equal, so that m = r for every i, then the common 
value is known as the radix of the system. In particular, the 
ordinary decimal number system is the radix 10 system, the 
binary system is the radix 2 system, and the Mayan system is the 
radix 20 system. For automatic calculating machinery, the use 
of a radix implies that the counters at each position can be identi- 
cal, so radical numbers are usually used; in cases such as the 
biquinary arithmetic described by Alt, 1 the use is essentially 
as decimal, or radix 10, arithmetic. 

For any radix arithmetic the basic tables corresponding to the 
addition and the multiplication tables of decimal arithmetic can 
be written, and from them the operations of addition, mul- 
tiplication, subtraction, and division can be carried out. 

For binary arithmetic, the tables are given in Tables 6-4 and 
Table 6-4. Addition Table 6-5. Multiplication 



~\^^ Augend 
Addend ^v^ 





1 




Sum 



1 



1 


1 
10 



\]VIultipli- 

^\cand 

Muliplier^\ 





1 




Product 




1 







1 



6-5. The rules of carry are the same as in arithmetic of any other 
radix. The construction of the tables is the same. 

For octal arithmetic, the tables are given in Tables 6-6 and 
Table 6-6. Addition 



v \ v ^ Augend 


















>v 





1 


2 


3 


4 


5 


6 


7 


Addend ^\ 
























1 


2 


3 


4 


5 


6 


7 


1 


1 


2 


3 


4 


5 


6 


7 


10 


2 


2 


3 


4 


5 


6 


7 


10 


11 


3 


3 


4 


5 


6 


7 


10 


11 


12 


4 


4 


5 


6 


7 


10 


11 


12 


13 


5 


5 


6 


7 


10 


11 


12 


13 


14 


6 


6 


7 


10 


11 


12 


13 


14 


15 


7 


7 


10 


11 


12 


13 


14 


15 


16 



ARITHMETIC SYSTEMS 



81 



Table 


6-7. 


Multiplication 






^\Multipli- 


















^\cand 





1 


2 


3 


4 


5 


6 


7 


Multiplier^ 











































1 





1 


2 


3 


4 


5 


6 


7 


2 





2 


4 


6 


10 


12 


14 


16 


3 





3 


6 


11 


14 


17 


22 


25 


4 





4 


10 


14 


20 


24 


30 


34 


5 





5 


12 


17 


24 


31 


36 


43 


6 





6 


14 


22 


30 


36 


44 


52 


7 





7 


16 


25 


34 


43 


52 


61 



6-7. These tables are unimportant for applications at present, 
but their structure should give a sufficient example to permit the 
reader to construct others as he may desire. 

Examples of binary and octal multiplication and division 
complete the exposition. To multiply 11,100,110 by 10,001,001, 
start the multiplication as always: 

11100110 
10001001 



11100110 
00000000 
00000000 
11100110 
00000000 
00000000 
00000000 
11100110 

In summing it is best to sum two nonzero numbers at a time until 
experience is gained. Thus the sum of the first and fourth rows 
from the top is 

11100110 
11100110 
100000010110 



where carries have occurred in all additions after the sixth. Then 
adding this sum to the last row, the final product is obtained 



82 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

100000010110 
11100110 



111101100010110 

with no carries. 

For an octal multiplication, consider multiplying 346 by 267. 
Again, the multiplying steps: 

346 
J7 
52 
34 
25 



346 


346 


6 


2 


44 


14 


30 


10 


22 


6 


2544 


714 


3112 




2544 




714 





3112 
The final totalization is 



122152 

Binary division is performed in much the same way as ordinary 
decimal long division. This may be illustrated in the following 
example by dividing the binary equivalent of decimal 137 by the 
binary equivalent of decimal 10: 

begi 
1101 



1010)10001001 
a 1010 
c 1110 
d 1010 



/ 10001 
h 1010 

j HI 

As shown above, the divisor is placed outside of the long-division 
sign, and the dividend is placed inside. By a series of successive 
subtractions, a quotient is formed above the long-division sign. 
The first step in the division process is to place the divisor beneath 
the dividend in a position as far to the left as a positive difference 
will allow. This is shown on line a. The first digit 1 of the 
quotient is placed at b in the same column as the lowest order 



ARITHMETIC SYSTEMS 83 

digit of the divisor. The first digit 1 of the divisor signifies 
that only one subtraction can be performed before a negative 
difference is encountered for this position of the divisor. The 
divisor is then subtracted from the dividend to produce the 
positive difference (line c). This difference is then compared 
with the divisor to note that it is smaller in value than the divisor. 
The next digit in the dividend is brought down to the difference 
obtained on line c. If the new number on line c is larger than the 
divisor, the divisor is then placed under this number and the 
subtraction process repeated. If the new number is less than 
the divisor, a is placed in the quotient in the same column as 
the lowest order digit of the divisor in this position. The next 
digit of the dividend is then brought down to the difference, 
the divisor shifted one place to the right, and the subtraction 
process is continued until the quotient is completed to the 
radical point. Any remainder after the last subtraction is 
treated in the same manner as in ordinary long division. In 
the example shown above, the quotient is 1,101 with a remainder 
of 111 -s- 1010; this corresponds to the decimal number 13^0- 

In the above example, it should be noted that no multiplication 
process was mentioned. This is apparent from the realization 
that the greatest number of allowable times the divisor goes 
into the dividend for any partial quotient is one, since this is the 
highest single-order digit in the binary system. Therefore, the 
product of the nonzero digit last added to the quotient and the 
divisor is the same as the divisor alone. This is not true, how- 
ever, for other radix systems. For example, use is made of the 
octal multiplication table, displayed above, in the performance 
of the same division in the octal system as follows: 

15 

12)211 

12_ 

71 

62 

7 

The quotient, as before, has a decimal equivalent of 13%o- 

6-6. A Note Concerning Binary -to -octal Conversion 

It is true according to Theorem 6-1 that exactly eight numbers 
can be represented with three binary digits. From this it follows 



84 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

that triplets of binary digits may be used as marks to represent 
octal digits, and that conversion from binary to octal notation 
simply implies changing the notation used for the marks. Spe- 
cifically, the notation change is as shown in Table 6-8. The 





Table 


6-8 




Binary Triplet 






Octal Mark 


000 









001 






1 


010 






2 


011 






3 


100 






4 


101 






5 


110 






6 


111 






7 



change is made triplet by triplet, beginning at the right. For 
example, the binary number 10,001,001 is the octal number 211; 
the binary 101,100,110 is octal 546, etc. 

The binary system has obvious engineering advantages in 
computer design. The expression of a number in binary notation 
requires a comparatively large number of digits, however. For 
example, a number with six decimal digits may require 20 binary 
digits; the same number, in octal notation, may need only seven. 

The number of octal digits is thus only slightly greater than 
the number of decimal digits, the conversion from binary repre- 
sentation is trivially easy, and an inexperienced reader can get 
good qualitative estimates of the results without worrying about 
the octal character of the representation. It can be seen, there- 
fore, that the octal system has certain features which make it 
very attractive. 

6-7. Economy Attained by Radix Choice 

The economy to be gained by choice of radix can be estimated 
as follows: From Theorem 6-1, it is known that the number of 
numbers expressible with n digits radix r is r n . It was noted in 
Chap. 3 that, using two-state tubes or relays, and for small 
values of r, each digit radix r requires r triodes or relays (but 
for larger values of r each digit requires up to 2r triodes or relays). 
Assuming that the complexity is measured according to the 
number of tubes, it is possible to determine the theoretical 
optimum value for the radix of the arithmetic system. 



ARITHMETIC SYSTEMS 85 

Let N = rn represent a fair estimate of the number of tubes 
required in the system and M = r n be the largest number of 
numbers expressible in the system, where r designates the radix 
and n designates the number of digits. The value of N should 
be a minimum, subject to M being fixed in value. Then 

M = r n (6-4) 

log M = n log a r = M f (6-5) 

M' 



IL — , 

loga r 
Substituting the value of n in the expression for N, 



(6-6) 



N 


loga 


r 
r 




g« r 


= log. 


r - loga 


e 


dN 


= M^ 


loga r — 


loga e) 


dr 


(loga 


r) 2 






M' 


^ 




(loga r) 


2 ^ U 




loga r 


- loga ( 


3 = 



(6-7) 
(6-8) 
= . (6-9) 

(6-10) 



Since 
then 

If 

and 

r = e = 2.71828 • • • (6-11) 

This is the only possible value for a minimum, and therefore N 
must be a monotonic function of r for r > e, and a monotonic 
function for r < e. 

Table 6-9 

r N 

2 39 . 20 

3 38 . 24 

4 39 . 20 

5 42 . 90 
10 60.00 

Now, consider some values, with M fixed at a value 10 6 , as 
shown in Table 6-9. Note that N increases with the values of r 
above 3. This increase must continue, for r has no further 
critical values. Under these assumptions, the radix 3, on the 
average, is the most economical choice, closely followed by radices 



86 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

Table 6-10 
Radix 



10 



Number Representation 



1 


1 


1 


1 


1 


1 


2 


2 


2 


2 


2 


10 


3 


3 


3 


3 


10 


11 


4 


4 


4 


10 


11 


100 


5 


5 


10 


11 


12 


101 


6 


6 


11 


12 


20 


110 


7 


7 


12 


13 


21 


111 


8 


10 


13 


20 


22 


1,000 


9 


11 


14 


21 


100 


1,001 


10 


12 


20 


22 


101 


1,010 


11 


13 


21 


23 


102 


1,011 


12 


14 


22 


30 


110 


1,100 


13 


15 


23 


31 


111 


1,101 


14 


16 


24 


32 


112 


1,110 


15 


17 


30 


33 


120 


1,111 


16 


20 


31 


100 


121 


10,000 


17 


21 


32 


101 


122 


10,001 


18 


22 


33 


102 


200 


10,010 


19 


23 


34 


103 


201 


10,011 


20 


24 


40 


110 


202 


10,100 


21 


25 


41 


111 


210 


10,101 


22 


26 


42 


112 


211 


10,110 


23 


27 


43 


113 


212 


10,111 


24 


30 


44 


120 


220 


11,000 


25 


31 


100 


121 


221 


11,001 


30 


36 


110 


132 


1,010 


11,110 


35 


43 


120 


203 


1,022 


100,011 


40 


50 


130 


220 


1,111 


101,000 


45 


55 


140 


231 


1,200 


101,101 


50 


62 


200 


302 


1,212 


110,010 


100 


144 


400 


1,210 


10,201 


1,100,100 


1,000 


1,750 


13,000 


33,220 


1,101,001 


1,111,101,000 


10,000 


23,420 


310,000 


2,130,100 


111,201,101 


10,011,100,010,000 



2 and 4. These assumptions are, of course, only approximately 
valid, and the choice of 2 as a radix is frequently justified on 
more complete analysis. It should be noted that, even with the 
optimistic assumption that 10 triodes will yield a reliable decimal 
ring, radix 10 leads to about one and one-half times the com- 



ARITHMETIC SYSTEMS 87 

plexity of radix 2, 3, or 4. This is probably significant despite 
the shallow nature of the argument used here. 

For reference, corresponding representations of the same 
numbers for radices 2, 3, 4, 5, 8, and 10 are tabulated in Table 
6-10. 

6-8. Negatives and Complements 

A desk calculator, in performing subtraction, uses a reversing 
gear and runs the wheels backward. If the subtraction goes far 
enough to generate a negative number, a difficulty is encountered. 
The difficulty, although easily resolved, arises because the 
indicated numbers continue to decrease, while in standard nota- 
tion they would suddenly be preceded by a minus sign and begin 
to increase. The solution lies in noting that the machine has a 
limited capacity. If the machine has 10-decimal-digit capacity 
in a register, it can hold numbers as high as 10 10 — 1, but it 
cannot hold 10 10 . The machine is unable to distinguish between 
two numbers which differ by any integral multiple of 10 10 ; in 
terms of the number theorist, it performs arithmetic modulo 10 10 . 
Any number is represented on this machine by the number plus 
or minus an integral multiple of 10 10 chosen in such a way as to 
bring the indicated number to a value between the limits 
and 10 10 — 1, inclusive. In particular, the number negative 137 
would be represented by -137 + 10 10 = 9,999,999,863. Once 
this condition is recognized, there is no difficulty if the coefficient 
of the additive 10 10 is kept in mind; for almost all calculations 
it is or 1. 

For many of the more rapid machines using other types of 
EPDCC's as computing elements, this system of representation 
of negative numbers and this system of subtraction are not satis- 
factory in general, since most of the faster EPDCC's cannot be 
made to run backward conveniently. Because such components 
will run in only one direction, the process of subtraction must be 
carried out in the manner defined in many algebra texts: change 
the sign and add. Furthermore, the question of changing the 
sign and adding cannot be treated too lightly, for if numbers are 
represented by their absolute values preceded by a plus or a 
minus sign, the result of adding a pair of Ts in the right-hand 
place depends on the signs of the numbers. Furthermore, if 
the radix is anything but 2, a binary mark must be used to denote 



88 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

the sign while all other marks are different from this. This 
situation may lead to unnecessary complexity in the machine. 

The number 9,999,999,863 written above is called the com- 
plement of the number 137 with respect to 10 10 . All the digits 
to the left of its last nonzero digit are obtained by subtracting 
the corresponding digits of the original number from 9 (with 
zero written in front as many times as is necessary). The last 
nonzero digit is obtained by subtracting the last nonzero digit 
of the original number from 10, and if there are any zeros to the 
right of this point in the original number, they are replaced by 
zeros in the complement. Thus the complement of 5,280 is 
9,999,994,720. The 9's complement of a decimal number is the 
number obtained by replacing each digit of the original number 
by 9 minus that digit To perform subtraction, a fast machine 
could form these complements and add, but this process is still 
inconvenient. Such a method demands that certain digits be 
treated in a special manner, and the positions of the specially 
treated digits vary from one number to another. 

The solution of treating all digits alike is so attractive that 
it has been tried arbitrarily, and it is used in most machines now 
being considered. The 9's complement of a positive number is 
used in this system to represent the negative number with the 
same absolute value. One disadvantage is immediately appar- 
ent: the 10-digit numbers 0,000,000,000 and 9,999,999,999 repre- 
sent the same number. This disadvantage is immediately offset 
by the fact that the devices in which these complements are to 
be used may add or they may subtract, but they cannot do both. 
The number 0,000,000,000 turns out never to be generated in a 
machine which always adds, and the number 9,999,999,999 is 
never generated in a machine which always subtracts. 

With 9's complements, numbers of 10 decimal digits are 
expressed modulo 10 10 — 1 rather than modulo 10 10 as above; 
i.e., numbers are stated with an implicit understanding that an 
integral multiple of 10 10 — 1 (with coefficient invariably or 1 
in common practice) has been added to the number originally 
to be represented. When two such numbers are added and their 
sum exceeds 10 10 — 1, the machine must automatically reduce 
them by this amount. To do this, it must detect the fact that 
the number is excessive and then correct. The following state- 
ments govern completely: 



ARITHMETIC SYSTEMS 89 

1. The sum of two 10-digit decimal numbers exceeds 10 10 — 1 
if and only if there is a carry beyond the 10th place. 

2. If there is a carry from the tenth place, the sum modulo 
10 10 — 1 may be represented with 10 decimal digits by 
adding this carry to the extreme right-hand digit. This 
is the so-called end-around carry. 

The first statement is obvious and requires no comment. The 
second is an automatic scheme for subtracting 10 10 from the sum, 
by not writing the carry in the 11th decimal position, and then 
adding 1 to the sum, by carrying to the first decimal place. The 
net result of this operation is to subtract 10 10 — 1; this is allow- 
able in the arithmetic modulo 10 10 — 1. 

Completely analogous rules hold for arithmetic with n digits 
radix r. They will be stated in a theorem the proof of which is 
omitted. 

Theorem 6-4. If x is a positive number expressed with n digits 
radix r, i.e., < x < r n — 1, then —x may be expressed modulo 
r n — 1 by replacing each of the n digits of x by r — 1 minus this 
digit; to add, modulo r n — 1, two numbers expressed with n digits 
radix r, proceed according to the rules of arithmetic radix r except 
that the carry from the nth digit, if it occurs, is to be added to the 
first digit (end-around carry). 

A particular example illustrates the principle. Two five-digit 
binary numbers will be added: 10,110 + 01,101. With ordinary 
binary arithmetic without end-around carry, the answer is 
100,011. To reduce to modulo 2 5 — 1, the sixth digit is deleted 
and added to the first with additional carries resulting to give 
00,100. 

In binary arithmetic of n digits it is usually assumed that 
numbers with the nth digit 1 are negative. Under this assump- 
tion the above addition becomes 1,101 — 1,001 = 100. 

6-9. Scale Factors and Radical Points 

The foregoing description of the arithmetic of integers in 
terms of radix r digits can be enlarged to include fractional and 
other numbers occurring in analysis. The basis of the enlarge- 
ment is classical, and only the most immediately applicable 
results will be given here. A complete description can be found 
in Ritt. 6 



90 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

Many of the numbers which are dealt with in scientific com- 
putation are derived from measurements of physical quantities. 
If such a quantity is to be represented in a register of n digits, 
it must be expressed in terms of a number of not more than n 
digits, whether or not it has been measured precisely enough 
to warrant expression in terms of a greater number. For 
example, to express w in a five-digit radix 10 register, we write 
3.1416. This expresses the fact that -k lies between 3.14155 and 
3.14165, which is entirely correct. 3.1416 is an arbitrarily pre- 
cise expression for a number the magnitude of which is known 
more precisely than can be expressed in five decimal digits. 

Furthermore, it is the use of the decimal point which permits 
this convenient representation of the number in five digits. 
3.1416 is a fraction. In terms of integers it is 31, 416/10 4 . The 
same scheme may be used with any other radix. If a radical 
point is placed to the left of the /cth digit (from the right) of a 
number expressed in radix r notation, the number is to be read 
as the integer, as if there had been no radical point, divided by 
r k . Thus 

07 _1_ 05 _!_ 03 _1_ 02 _|_ 1 

Binary 101.01101 = decimal — — L 

= decimal 17 %2 
= decimal 5.406 

Note that the numerator and the implied denominator must 
both be translated if the radix is changed and that the rule in 
Theorem 6-3 cannot be carelessly applied to this case of frac- 
tional representation. It is not difficult, however, to devise 
a multiplication rule similar to the division rule of Theorem 6-3 to 
handle that portion of the number to the right of the radical 
point. 

For machines with the radical point in a fixed position, and 
these machines are the simplest to build, the use of scale factors is 
necessary to obtain the most efficient utilization of the limited 
number of numbers representable by the machine. For this, 
the units in which the numbers appear are measured and rechosen 
in such a way that the largest number of any kind which will 
appear in the machine is just slightly smaller than the largest 
number the machine will hold. This is an operation analogous 
to measuring in inches rather than feet. 



ARITHMETIC SYSTEMS 91 

Some workers feel that there is a best place for the radical 
point. In particular, many, including Burks, Goldstine, and 
von Neumann, 2 feel that this best place is one binary digit to 
the right of the left-hand end of a binary number. It should be 
noted that the first left-hand digit does not represent a numerical 
value but merely serves to indicate whether the number is posi- 
tive or negative. This placement used with l's complements 
gives numbers ranging from —1 to +1, and the product of two 
such numbers lies in the same range. This feature, which does 
not demand a scale change when the products are formed, is 
one of the principal advantages of this system. 

Other workers have arbitrarily placed the radical point at the 
right-hand end of the number; i.e., they have built their machines 
to deal with integers. Thus the basic operations are programmed 
on the machine in terms of shifting operations. One such machine 
has been described by Engineering Research Associates, Inc. 

A third possible scheme is one in which the position of the 
radical point is not designated in advance but is placed by the 
machine in performing each operation. This leads to engineer- 
ing complexities, but it serves to reduce the burden of converting 
all numbers involved into a scale best suited for the machine. 
The Mark II Calculator built at the Computation Laboratory 
at Harvard University for the Naval Proving Ground at Dahl- 
gren, Va., is an example of such a machine. 

6-10. Binary Division — A Special Case 

Binary division offers simplifications which have been noted 
by Burks, Goldstine, and von Neumann. 2 These authors sug- 
gest that the division process for binary numbers be modified to 
proceed according to the following steps in a system of commands: 

1. Enter the dividend and the divisor so that like order digits 
are in the same columns. 

2. Shift the divisor to place its first left-hand nonzero digit 
in the same column as that of the dividend; add as many 
0's at the right-hand end of the divisor as necessary to 
make it equal in order to the dividend. 

3. Subtract the divisor from the dividend and register a 1 
in the first-order-digit position of the partial quotient 
register. To establish its proper position, shift this 1 a 



92 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

number of digit places, to the right or to the left, cor- 
responding to the number of places and direction by which 
the divisor was shifted in step 2. 

4. For the next arithmetic operation, shift the divisor one 
place to the right. 

5. If the remainder (difference) after step 3 is positive, sub- 
tract the shifted divisor from this remainder and add 1 to 
the partial quotient in the next position to the right of the 
previous 1; if the remainder is negative, add the shifted 
divisor to it and subtract 1 from the partial quotient in the 
new position. 

6. Shift the divisor and proceed as in step 4 using the new 
remainder and modifying the quotient by 1 in the next 
position. 

7. Continue until the process of division is completed, at this 
time adding the finally shifted divisor one additional time 
in carrying out a corresponding subtraction from the quo- 
tient, if necessary, to get a positive remainder. 

As an example, a division which occurs as the first step in 
translating binary 10,001,001 to a decimal number will be per- 
formed, keeping all arithmetic in the binary system and using 
Theorem 6-3. It should be noted that where a negative differ- 
ence is obtained in Table 6-11, it is represented by its l's com- 
plement plus 1. 

Since the last sum in Table 6-11 is positive, the division is 
complete. The quotient is 1,101 (decimal 13), and the remainder 
is 111 (decimal 7); these numbers agree with those found earlier. 

The justification of this method, mathematically, lies in the 
trivial proof that the dividend is the sum of remainder and the 
product of the quotient with the divisor. In terms of machine 
operation the justification lies in the ease of programming and 
in the computing speed. The example is included as an appli- 
cable division scheme for binary digital computers. 

The applicability of this scheme is enhanced by noting that a 
simple rule may be formulated for carrying out the division 
without borrows in the quotient. Consider that the quotient 
is formed from two numbers x and y such that the quotient z is 
x — y and such that x + y, in binary notation, is a series of l's. 
That is, the quotient 



with 



ARITHMETIC SYSTEMS 
z = x — y 
x + y = 2 n - 1 



for some n. It follows from the second expression that 

z = 2x - 2 n + 1 
In the above example, 

z = 1,101] 
x = 10,110> 



93 

(6-12) 

(6-13) 
(6-14) 

(6-15) 



y 



1,001 



where x is formed with l's in the places where the divisor was 
added, and y is formed with l's in the places where the divisor 

Table 6-11 







Step 








(Add next 




Operations 




step if 

minus; 

subtract 

if plus) 


Partial quotient 


(a) 








Dividend (decimal 137) 


10001001 






Divisor (decimal 10) 


1010 






Difference 


-010111 


- 


10000 


(b) 




/ 




Divisor (shifted) 


1010 


r 




Algebraic Sum 


111001 ' 


+ 


10000 - 1000 = 1000 


(c) 




/ 




Divisor (shifted) 


1010 y 


* 




Difference 


10001 / 


+ 


1000 + 100 = 1100 


(d) 




y 




Divisor (shifted) 


1010 > 


* 




Difference 


-11 X 


— 


1100 + 10 = 1110 


(e) 




y 




Divisor (shifted) 


1010 j 


r 




Algebraic Sum 


111' 


+ 


1110 - 1 = 1101 



was subtracted. The operation defined in Eq. (6-14) is brought 
about by shifting all digits one place to the left, adding a at 
the end (multiplication by 2), deleting the initial 1 (subtraction 
of 2 n ), and then replacing the final by a 1 (addition of 1). 



94 THE BASIC ELEMENTS OF MACHINE COMPUTATION 
Thus, in the example, step by step, 

x = 10,110 -> 101,100 -» 01,100 -> 01,101 -> 1,101 = z 
The negative part of the quotient need never enter the calculation. 

6-11. Alternate Arrangements — Linear Mappings 

The system of complements described in Sec. 6-8 above is 
an efficient and convenient method for handling numbers which 
occur in calculations where the range of numbers extends approxi- 
mately equally on both sides of zero. The advantages are 
twofold: 

1. Registers with n radix r positions can be utilized fully 
without use of additional sign symbols. 

2. The arithmetic steps are carried out digit by digit as they 
are in long division, long multiplication, addition, and 
subtraction with ordinary numbers; the behavior of each 
digit is independent both of the value of other digits and the 
sign of the numbers. 

In the addition process, the system of arithmetic modulo r n — 1 
has the advantage that complements are easily formed and 
arithmetic can proceed with either an addition or a subtraction 
process. The first of the two properties of the complement 
system displayed above can be shown to be shared by linear 
mappings, and except for some permutations of digits which do 
not essentially change the results, these are the only mappings 
which have the advantage of permitting digit-by-digit calcula- 
tion. The linear mapping is defined as follows: 

Definition: Let the number x, < x < r n — 1, be an integer 
representing the ordinal number of any of r n states assumed by an 
n-digit radix r register and let a(^ 0) and b be two constants. 
Then, under a linear mapping, the coefficients a and b and the 
state x represent the number y with 

y = ax + b (6-16) 

Under this definition it can be shown easily that multiplication of 
two numbers y± and y 2 can be brought about by operating, digit 
by digit, on the numbers Xi and x 2 which represent them. Indeed, 
if 

2/i = a Xl + b (6-17) 



ARITHMETIC SYSTEMS 95 



and 










?/2 = ax 2 + b 




(6-18) 


and if 










ViVi = y* = ax 3 + b 




(6-19) 


then 










b 2 - 
xz = axix 2 + b(xi + x*) + 


- b 


(6-20) 



With a and b constant, this expression can be built up under 
fixed rule, digit by digit, just as the product in ordinary arith- 
metic is built up. Similarly, the operations of addition, subtrac- 
tion, and division may be carried out on a digit-by-digit basis. 

Now, if the y values range between a and (3 so that a < y < /3, 
then it is possible to position the numbers a and b in such a way 
that a corresponds to the value x --= and corresponds to the 
value r n — 1. This possibility permits the use of the linear 
mapping system to give the registers the greatest possible effi- 
ciency. To obtain this result, it is only necessary to write 

b = a 

- a (6-21) 



a 



r n — 1 



Some simple examples of linear mappings are listed here. In 
the variant of binary arithmetic described in Sec. 6-10, the 
representation of the quotient z by the positive part of the 
quotient x was defined in Eq. (6-14) ; examination of this equation 
shows that it is a linear mapping with a = 2 and b = — 2 n + 1. 
A linear mapping with piecewise constant coefficients can be 
used to define systems of complements to give arithmetic modulo 
r n and modulo r n — 1; in either case, for y nonnegative, the 
coefficients are a = 1 and b = 0; in the first case, for y negative, 
the coefficients are a = + 1 and b = — r n , and in the second case 
the coefficients are a = +1 and b = 1 — r n . 

In general, the full utilization of the range of y numbers 
appearing will not be the only consideration in choice of a and b. 
Among other factors which should be considered are ease of 
subtraction, elimination of round-off bias, and decrease of 
number of trial subtractions in division. 

6-12. Note on Round-off Error 

In this chapter the elements of the arithmetic theory useful 
in applying EPDCC's to computing problems have been out- 



96 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

lined. The chapter has shown how by proper choice of radix 
an EPDCC with any number of stable states can be applied to a 
computing problem. 

It must be remembered, however, that many other details must 
be taken into account. Not only are engineering details involved 
in connecting the components to carry out the calculation of 
crucial importance, but also other mathematical details must be 
considered. In particular, the extremely fast calculating speed 
of these machines introduces a new possibility of round-off 
accumulating. Although the machines function in the same 
way that a desk machine functions, it is no longer feasible to 
overcome the round-off error by sheer length of numbers. The 
way in which these errors accumulate might be justifiably 
included in this chapter; however, the application of methods for 
estimating the accumulation of round-off errors to problems of 
numerical calculation is involved, and its exposition is likely to 
be long and understood only by computers willing to devote 
considerable time to the study. A detailed treatment of round-off 
error is given by Burks, Goldstine, and von Neumann 2 and by 
von Neumann and Goldstine. 5 

6-13. Appendix 

6-13-1. Proof of Theorem 6-1. The number of different 
numbers which can be represented by configurations with marks 
differing from in at most the first k positions is the product 

P k = nin 2 n 3 • • • n k (6-1) 

Proof: It is a classical proposition that Pk different configura- 
tions can be made with the k marks in order, and it is obvious 
from the definition of counting (meaning exactly that it can be 
proved by mathematical induction) that no two of these con- 
figurations represent the same number. This completes the 
proof. 

6-13-2. Proof of Theorem 6-2. In the representation of a 
number x, if pi represents the number of marks admissible which 
precede the mark actually occurring in the first position, if pi is 
the number of marks admissible for the second position which 
precede the mark in that position, etc., with pi the number of admis- 
sible marks for the ith position preceeding the ith mark, then, 



ARITHMETIC SYSTEMS 97 

conforming with the notation for P of Eq. (6-1), 

x = pi + V2P1 + P3P2 + • ' ' + ViPi-i (6-2) 

Proof: The proof of this theorem will be by mathematical 
induction. It will be shown that the set of values of x for which 
the statement is true includes the number zero and that if this 
set includes all nonnegative numbers less than a number N, it 
also includes N. All such sets clearly (or axiomatically) include 
all the nonnegative integers, and it will follow that the theorem 
is true — that the statement is true for every value of x. 

First, it is clear that if x = 0, the statement is true, for then 
by definition all the numbers p t - are zero, and the right member 
of Eq. (6-2) is zero and therefore equal to x. Now, suppose that 
when x is equal to some particular number N, Eq. (6-2) holds. 
Then if the representation of the number N + 1 is written, one 
or more of the numbers jpi determined by this new number differ 
from the corresponding numbers in the representation of N. 
According to the counting rule, if p k is the last one which differs, 
then it is greater by 1 in the representation of N + 1 than it 
was in the representation of N; furthermore, since all pi with 
i < k must have been maximal in the representation of N and 
since they have all been replaced by zero, each such p, in the 
representation of JV + 1 is n,- — 1 less than it was in the repre- 
sentation of N. Now, it was assumed that Eq. (6-2) was valid 
when x was N. When x is replaced by N + 1, it is clear that the 
left member is increased by 1, and it has just been remarked that 
the right member is increased by the quantity 

Q k = Pk-i - (nw - l)P*-2 -.■••- (n, - 1) (6-22) 

The theorem will have been established if it can be shown that Q k 
is always 1 no matter what value k takes. [For example, for 
decimal numbers it is true that 

Q k = io*-i - 9(10 fc - 2 + 10 fc - 3 + • • • + 1) (6-23) 

and this expression is readily seen to be exactly 1 no matter what 
value k may have by the theorem giving the sum of terms in a 
geometric series.] The fact that Q k is always 1 can be established 
rigorously by mathematical induction or more intuitively by 
noting that P k -\ is equal to n k -\Pk-2 by definition and that the 
various terms of the right member of Eq. (6-22) annihilate one 



98 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

another in succession until only the final number 1 is left. This 
completes the proof of Theorem 6-2. 

6-13-3. Proof of Theorem 6-3. If x is a nonnegative integer 
and if ?u, pi, and Pi all have the significance defined above for some 
system of notation in which x is expressed, then p\ is the remainder 
when x is divided by n\, p 2 is the remainder when this integral 
quotient (without remainder) is divided by n 2 , etc. : explicitly, there 
exists a set of nonnegative integers q , q h q 2 , . . . , qi such that 

q Q = x = ftitfi + Pi', = pini 
qi = n 2 q 2 + P2) = p 2 n 2 
(6-3) 

q { = n 1+1 q< +1 + p i+1 ; = p i+ in i+ i 



These equations uniquely determine pi. 

Proof: It is clear that Eqs. (6-3) define the p's, for these equa- 
tions are only a formal method of stating that, say, qi is the 
quotient and pi the remainder when qi_i is divided by n». With 
this in mind, the following inductive proof will make use of an 
expression which will be shown for every value of i to define qi 
equivalent to the definition in Eq. (6-3) subject to the original 
restriction q = x: 

q% = Vi+i + Pi+2n i+ i + pi + 3n i+2 n i+ i + p i+4 n i+3 n i+2 ni +1 + • • • , 

i = 0, 1, • • • (6-24) 

To establish this for i = 0, all that is required is to replace i by 
throughout and to compare the result with Eq. (6-2), which has 
already been established; if the P's in Eq. (6-2) are written in 
terms of the n's, the two will be identical. Next seek to deter- 
mine whether the set of values of i, for which Eq. (6-24) gives 
the same value of qi as Eqs. (6-3), does possess the property that 
the set contains the next higher integer to any integer it contains. 
Suppose that for i = k, the result from Eq. (6-24) is the same as 
that from Eqs. (6-3). Then, for this value of k, 

qk = Pk+i + Pk+zrik+i + p k +zn k+2 n k+ i + • • ■ (6-25) 

Rewritten, this states 

q k = n k+1 (p k+2 + p k+3 n k+2 +•••) + Pk+i (6-26) 



ARITHMETIC SYSTEMS 99 

Now, according to Eqs. (6-3) with i = k, this should be 

qk = n k+1 q k+1 + p k +i (6-27) 

and the expression in the parentheses, which therefore must be 
qk+i, is indeed the value that Eq. (6-24) gives for q k+ i by setting 
i = k + 1. Thus, whenever Eq. (6-24) gives the right answer 
for any value of i (as it does for i = 0), it also gives the right 
value for the next higher value, and the statement of the theorem 
is true for these values. This completes a proof by mathe- 
matical induction, for the set of values of i for which the state- 
ment of the theorem is true must contain all positive integers 
if it is to fulfill these requirements. 

REFERENCES 

1. Alt, Franz, A Bell Telephone Laboratories' Computing Machine, Mathe- 
matical Tables and Other Aids to Computation, Vol. Ill, pp. 1-13, 69-84 
(1948). 

2. Burks, A. W., H. H. Goldstine, and John von Neumann, Preliminary 
Discussion of the Logical Design of an Electronic Computing Instrument, 
Part I, 2d ed. (Institute for Advanced Study, Princeton, N.J., 1947), 
pp. 23-29. 

3. Carmichael, Robert D., The Theory of Numbers, (John Wiley & Sons, Inc., 
New York, 1914), pp. 22-24. 

4. Morley, S. G., An Introduction to the Study of the Mayan Hieroglyphs, 
Bulletin 57, Bureau of American Ethnology (The Smithsonian Institution, 
Washington, D.C., 1915). 

5. von Neumann, John, and H. H. Goldstine, Numerical Inverting of 
Matrices of High Order, Bulletin of the American Mathematical Society, 
Vol. 53, pp. 1021-1099 (1947). 

6. Ritt, J. F., Theory of Functions (King's Crown Press, New York, 1947), 
pp. 1-10. 

7. Spinden, H. J., Ancient Civilizations of Mexico and Central America, 
Handbook Series No. 3, 3d ed. (American Museum of Natural History, 
New York, 1928), pp. 118-121. 



CHAPTER 7 
NUMERICAL ANALYSIS 



7-1. Introduction 



In this chapter attention will be given to the numerical methods 
for solving certain of the principal types of equations dealing 
with algebraic, differential, and integral elements. As stated 
in the introductory chapter, these methods are regarded as 
components in the art of computation to the same degree as the 
physical equipment described in Chaps. 3, 4, 13, and 17. The 
material dealing with methods of computation for algebraic 
and differential equations is so extensive that the scope of this 
chapter will be limited to a brief outline and to the discussion of a 
few methods which are applicable to the principal types of equa- 
tions to which a digital computing machine is expected to be 
applied. For the sake of simplicity, this chapter is organized to 
treat first with the process of interpolation and interpolation 
formulas since this procedure forms the basis of many numerical 
methods. There will then follow brief outlines of methods for 
numerical differentiation and integration which depend largely 
upon these interpolation formulas. The numerical solution of 
algebraic and transcendental equations will next be taken up. 
Finally a discussion will be given concerning the procedures for 
numerical solution of total differential equations. 

The material for this chapter has been drawn largely from such 
standard texts as Ford, 7 von Karman and Biot, 12 Margenau and 
Murphy, 15 Scarborough, 19 and Whittaker and Robinson, 20 and 
from such special reports as Bennett, Milne, and Bateman 3 
and Bargmann, Montgomery, and von Neumann. 2 

7-2. The Process of Interpolation 

The process of interpolation consists essentially in representing 
a function in a particular form by making use of values which this 
function assumes for certain given values of the argument, or the 
independent variable. For example, let y = f(x) take on values 

100 



NUMERICAL ANALYSIS 101 

2/o, yi, 2/2, • • • , Vn for respective values of the argument x , Xi, 
Xtj ... , x n . Now let 4>{x) represent some arbitrary and, 
perhaps, simpler function such that it takes on the values of f(x) 
for the same values of the argument x , xi, x 2 , . . . , x n . If 
f{x) is now replaced by cf>(x) over the interval of the independent 
variable, then this procedure is known as interpolation, and the 
function 4>{x) is called the formula of interpolation. The mathe- 
matical basis for this procedure stems from two powerful theorems 
by Weierstrass. These theorems may be stated as follows: 

1. Every function continuous in the interval (a,b) may be 
represented in that interval, to any degree of accuracy, by a 
polynomial P(x) such that the absolute value of the differ- 
ence between the function and this polynomial may be made 
smaller than any previously assigned positive quantity e. 

2. Every continuous function periodic in 2ir may be repre- 
sented by a finite trigonometric series 

F(x) = a + ai sin x + a 2 sin 2x + • • • + a n sin nx 

+ 6i cos x + b 2 cos 2x + * * * + b n cos nx 

where the absolute value of the difference of the function 
and this trigonometric series may be made smaller than any 
previously assigned positive quantity 8. 

Let us consider a function y = f(x) for which 10 values of the 
function are known for 10 values of the independent variable. 
It will now be instructive to form what is known as a difference 
table for these values of the function and the independent variable. 
We arrange the values of the independent variable and of the 
function in the first two columns, respectively, and proceed to 
take differences between these values of the function, where 
As/o = 2/i - 2/o, A?/i = 2/2 - 2/i, ... , Ay n = y n+1 - y n ; these 
are denoted as first differences and appear in the third column. 
Second, third, fourth, and higher differences up to ninth differ- 
ences may be obtained in the same manner and give rise to the 
elements appearing in the diagonal difference table (Table 7-1). 

Such an arrangement is important because it may be shown 
(1) that if the values of the independent variable are taken in 
arithmetic progression, the nth differences of a polynomial of the 
nth degree are constant, and (2) that if the nth differences of a 
function are constant for an arithmetic progression of the 



102 THE BASIC ELEMENTS OF MACHINE COMPUTATION 



argument of the independent variable, this function is a poly- 
nomial of degree n. The latter means, essentially, that if an 
unknown function y = f(x) is tabulated in a difference table 
and if the fourth differences, for example, are found to be con- 
stant, then such a function may be approximated by a fourth- 
degree polynomial <$>(x) within the region of the limits of data 
given for the independent variable. 

Table 7-1 



X 


y 


Ay 


A 2 2/ 


A 3 ?/ 


A 4 2/ 


A 5 !/ 


A 6 2/ 


A 7 z/ 


A s y 


A 9 !/ 


X 


2/o 


A2/o 


















Xx 


2/i 


As/i 


A 2 2/o 


A 3 2/o 














x 2 


2/2 


A2/2 


A 2 2/i 


A 3 ?/i 


A 4 2/o 


A 5 2/o 










x 3 


yz 


A?/3 


Ahj 2 


A 3 2/2 


A 4 2/i 


A 5 Z/i 


A 6 2/o 


A 7 2/o 






Xi 


y* 


A2/4 


A 2 Z/3 


A 3 2/s 


A 4 2/ 2 


A«yi 


A 6 2/i 


A 7 ^! 


A 8 2/o 


A 9 2/o 


x 5 


y& 


A2/5 


A 2 l/4 


A 3 2/4 


A 4 ?,' 3 


A 5 l/3 


A 6 2/ 2 


A 7 2/2 


A 8 ?/i 




Xe 


2/6 


Aj/e 


A 2 l/5 


A 3 ?/ 5 


A 4 2/ 4 


A 5 2/4 


A 6 2/3 








x 7 


2/7 


A2/7 


A 2 2/e 


A 3 ?/6 


A 4 2/ 5 












Xs 


2/8 


Ay 8 


A 2 2/7 
















Xq 


2/9 





















where 



and 



At/o =2/i-2/o 

A?/i =2/2-2/1 

At/2 =2/3-2/2 

A?/ n = i/ n+1 - 2/*, 



A 2 2/o = Ay/i - At/o = 2/2 - 2j/i + 2/o 
A 2 2/i = A2/2 - A2/1 = 2/3 - 22/2 + 2/1 

A 2 2/n = A?/ n+1 - A2/ ra = ijn+2 - 2y n+1 + J/n 

The application of machines for tabulating differences of a 
function has recently been described by Laderman and Abramo- 
witz. 13 The Underwood-Elliott Fisher Sundstrand Accounting 
Machine, Model D, has been found exceptionally efficient for 
the construction of difference tables and for computing and 



NUMERICAL ANALYSIS 103 

printing the values of a function when initial differences and 
values of the function are given. 

7-2-1. Newton's Interpolation Formula. It will be instructive 
here to show how a polynomial of degree n may be made to 
approximate a given function f(x) in the interval defined by the 
argument. There are a number of interpolation formulas which 
may be used to define a function <f>(x) as an approximation to a 
given function /(x). Since the derivation of all these formulas 
is substantially similar, however, Newton's formula for forward 
interpolation will be used here as an example. Several remarks 
will be made concerning the application of this formula and 
several other related formulas of interpolation. 

Let the function y = f(x) take on values y , y 1} y 2) . . . , y n 
for respective values of the independent variable x , Xi, x 2 , . . . , 
x n . We are now interested in finding that polynomial <f>(x) which 
is required to represent f(x) in the interval determined by the 
range of the independent variable, whose successive values are 
taken as equidistant. The polynomial <f>(x) of the nth degree 
may be written as follows: 

<j>(x) = do + di(x — xo) + a 2 (x — x )(x — x\) 
+ a s (x — xo)(x — Xi)(x — x 2 ) 
+ a 4 (x — xo)(x — xi)(x — x 2 )(x — x s ) 
+ • • • + a n (x - x )(x — xi)(x — x 2 ) ' • • (x — x n -i) (7-1) 

In order to make <^(x ) = y , 4>(xi) = y h . . . , <f>(x n ) = y n , it is 
necessary to determine the coefficients a , ai, a 2 , . . . , a n . 
This may be done by substituting the successive values x , x h 
x 2 , . . . , x n into the relation for <f>(x) above, with the added 
requirement that the value of the arithmetic progression in the 
independent variable be taken as h such that x\ — x = h, 
x 2 — Xo = 2h, . . . , x n — Xo = nh. With these substitutions 
the polynomial 4>{x), known as Newton's formula for forward 
interpolation, may be derived; this is written as follows: 

4>i(x) = 



■yo + x (x 


- x ) -f 


2/i 2 {X 


— x )(x 


- Xi) 




X )(X — 


xi)(x — 


x 2 ) 






Xo)(x — 


xi)(x - 


x 2 )(x — 


Xz) + 


+ ^(x- 


Xo)(x — 


xi)(x - 


x 2 ) • • 


. (*- 



Xn-0 (7-2) 



104 THE BASIC ELEMENTS OF MACHINE COMPUTATION 
This may be simplified by a change of variable such that 



x — Xo 

; = U 



From this relation it follows that (x — Xi)/h = u — 1, etc.; the 
generating relationship 19 is then seen to be 

= u — n + 1 



h 

Newton's formula for forward interpolation with this change of 
variable takes the form 

<h(x) = 2/o + u Ay + -A_ — I A 2 ?/o + 3! A 3 y 

u(u - l)(u - 2){u - 3) 
+ 4j A 4 2/ 

+ ■ - • + <u-l){u-2)^ ■ (u-n + 1) ^ (7 _ 3) 

The forward-interpolation formula of Newton is so named because 
this formula treats values of the tabulated function starting 
with y and proceeding forward to y n ; it does not allow inter- 
polation below or to the left of y . For this purpose there exists 
another form of this formula known as Newton's formula for 
backward interpolation: 

J. ( \ I A I U ( U + 1) A 

4>ii(:r) = y n + u Ai?/ n H ^ A 22/« 

+ u(« + IKu + 2) Azyn + u(« + 1)(, + 2)(, + 3) Atyn 

+ ■■■ + u(u+l)(u + 2)y(u + n-l) Kyn (7 _ 4) 

where Ai2/ n = y n — y n -\ 

A 2 ?/n = yn- 2y n -i + 2/«-2 



7-2-2. Stirling's Interpolation Formula. In the previous 
section, Newton's formula for forward interpolation 4>i(x) was 
related to the array of differences in the diagonal difference table 
(Table 7-1). It was seen there that this interpolation formula 
was applicable only to determining the difference of values of 
the function y = f{x) for a forward progression y , y h y 2} ... , y n . 
Similarly the Newton formula for backward interpolation allows 



NUMERICAL ANALYSIS 



105 



differences to be determined for the progression y n , y n -i, y n -2 t 
. . . , 2/1. These formulas are fundamental to the interpolation 
process; in fact, other interpolation formulas for functions with 
equidistant increments of the argument may be derived directly 
from Newton's formulas by simple algebraic transformations. 
The only probable disadvantage in the use of Newton's formulas 
is that they quite often do not converge rapidly enough for certain 
practical applications. Because they obviate this difficulty, 
interpolation formulas due to Stirling and to Bessel, based on a 
slight variation of the original tabulation of differences, are 
important; these may be applied in cases where Newton's 
formulas do not provide rapid enough convergence. 

Let us return to a function y = fix) with 10 values of the 
function given for 10 values of the argument. We now arrange 
these values of the function in a table where the designation of 
the 2/0 position lies at or near the middle of the first column. 
This is known as a central-difference table (Table 7-2). 



Table 7-2 



X 


V 


Ay 


A*y 


A 3 y 


A 4 ?/ 


A 5 y 


A 6 y 


A 7 |/ 


A 8 2/ 


A 9 ?/ 


Z_4 
X-3 


V-4 

y-z 
y-2 
y-i 


Ai/-4 

Ays 

&y-2 


A 2 2/_4 

A 2 7/-3 

A 2 ?/- 2 


A 3 2/_4 
A 3 t/_ 3 


A 4 ?/_4 
A 4 ?/-3 


A 5 2/-4 


A 6 ?/-4 








X 


yo 


A?/o 


A 2 ?/-! 


A 3 2/- 2 
A 3 ?/_i 


A 4 2/_ 2 


A 6 2/_3 
A 5 y_ 2 


A 6 Z/_3 


AV4 
A 7 JM 


A 8 ?/_4 


A 9 ?/-4 


Xi 

x 2 
x z 

Xi 

x h 


y\ 

2/2 

y* 
y* 
y& 


&yi 

A2/2 

&V4 


A*y 

Ah Jx 

A 2 ?/2 
A 2 T/3 


A 3 2/o 
A 3 |/i 
A 3 y 2 


A 4 2/_i 

AV 

AVi 


A 5 t/_i 
Ahjo 


A 6 y_ 2 
A 6 r/-i 


A 7 2/- 2 


A 8 ?/-3 





In this table only the differences lying across the table immedi- 
ately above and below 2/0 (marked by broken lines) are used in 



106 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

the interpolation formulas of both Stirling and Bessel. Both 
Stirling's and Bessel's formulas may be derived from a central- 
difference table; it is found slightly easier, however, to obtain 
these two formulas by an algebraic transformation of the coeffi- 
cients of differences appearing in Newton's formula for forward 
interpolation [Eq. (7-3)]. It is seen that the coefficients in u 
of the various orders of differences in yo are actually the binomial 
coefficients. 19 By algebraic transformation these coefficients 
in u may be replaced by the central differences to which we have 
previously referred. By this procedure, we obtain Stirling's 
formula for interpolation: 

. A?/_i + A?/ . u\ 2 u(u 2 - l 2 ) A 3 ?/_ 2 + A z y-i 

, ^ 2 2 - l 2 ) w i u ( u2 - 12 )0 2 - 2 2 ) A 5 2/_ 3 + A b y-2 
H -^n A V-2 + 



4! *-* ' 5! 

u(u- l)(u-2) Afi _. , 



u(u 2 - 


6! 

- V)(u 2 - 


- 2 2 )(u 2 - 3 2 ) • • • [u 2 - (n - l) 2 ] 


u 2 (u 2 


- \ 2 ){u 2 - 


(2n - 1)! 

^ A 2n ' l y- n + A>*-Hr-(— d 
2 
- 2 2 )(u 2 - 3 2 ) • • • [u 2 - (n - l) 2 ] 






(2ti)! 

X A 2 ^_ n (7-5) 



+ 
+ 

+ 



7-2-3. Bessel's Interpolation Formula. By a similar pro- 
cedure, using another substitution for the binomial coefficients 
in u appearing in Newton's formula for forward interpolation, we 
obtain BesseVs formula for interpolation: 

y = V-l + V} + ( w - 1) Ayo + ^J) AV. + A«y. 

(m - ^)m(m - 1) A . , m( m - 1)(m + 1)(m - 2) 
3! ^ tf-i -r 4 | 

A 4 ?/_ 2 + A V-! 

X "" 2 

_l (u - HM* " l)(t* + l)(n - 2) 

n 2/_2 



+ 



5! 
w(t* - l)(u + l)(w - 2){u + 2)(w - 3) 



6! 

A 6 t/_ 3 + A 6 y_ 2 , 



+ 



NUMERICAL ANALYSIS 107 

u(u - 1)Q + 1)0 - 2)Q + 2) • • • Q - n)(u + n - 1) 

(2n)\ 

X " 



2 

+ O - 3^)w(w - 1)0 + 1)0 - 2)0 + 2) • • • 

(u — n)(u + n — 1) 

(2n + 1)! 

X A 2n+1 y- n (7-6) 

7-2-4. Lagrange's Formula of Interpolation. The formulas of 
Newton, Stirling, and Bessel apply only when equidistant values 
of the argument are given. Lagrange, however, developed an 
interpolation formula which is applicable to a function for which 
the values of the argument are not equidistant. 

Consider the function y = f(x) for which it is desired to obtain 
an interpolation polynomial <f>{x) of the nth degree. If this 
nth-degree polynomial is written as 

<f>(x) = a (x — Xi)(x — x 2 )(x — Xz) - • - (x — x n ) 

+ ai(x — x )(x — x 2 ){x — x 3 ) • ' - (x — x n ) 
+ a 2 (x — Xo)(x — xi)(x — x 3 ) • ' - (x — x n ) + • • • 

+ a n (x — Xo)(x — xi){x — x 2 ) • ' ' (x — Xn-i) (7-7) 

the n + 1 constants a h a 2 , . . . , a n must now be determined in 
order that <£(a;o) = Vo, <^(^i) = Vi, • • • , <t>(x n ) = y n . By 
taking x = x and <t>(x ) = y , a may be evaluated. By a 
similar procedure using X\ and 4>(xi) — y h ai may be evaluated, 
etc., up to and including the evaluation of a n . By replacing the 
values for the n -\- 1 constants by their evaluated operations, 
Lagrange' s formula of interpolation is obtained: 

<h( x ) = Qe - xi)(x - x 2 ) • • • (x - x n ) 

(Xq — Xi)(^o — X%) ' ' ' (Xq — X n ) 



+ 



(X — X )(X — X 2 ) • • • (X — X n ) 
(Xi — X )(Xi — X 2 ) ' ' • (Xi — X n ) 



, (X — X )(X — X^)(X — X Z ) ' ' ' {X — X n ) 

(X 2 — X )(X2 — Xi){x 2 — X 3 ) * ' ' (X 2 — X n ) 

i ... i \ x ~ x o)\ x ~ X U ' ' ' {x — X n -\) C7-9C) 

\X n Xo){x n X\) ' ' ' [X n X n —\) 

Since Eq. (7-8) is a functional relation between the variable y 
and the variable x, it is allowable to interchange the roles of the 
dependent and independent variables and obtain a functional 



108 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

relation between y and the various given values of x as follows: 

e(v) = ^ y ~ Vl ^ y ~ V2) ' ' ' (y ~ Vn) x Q 
(Vo - 2/0 (yo - 2/2) • • • (2/0 - y») 

, (y - yo)(y - y») • • • (y - y») ^ 
(yi - yo)(yi - yi) • ■ ■ (yi - y») 

_i_ (y - yo)(y ~ yi)(v - ys) • • • (y - y») ^ 2 
(ya - ye)(y2 - y0(y2 - y 3 ) • ■ • (y 2 - y») 
4. . . . + ^ ~ J^Kff - yO • • • (y - yn-0 x (7 _ Q) 
(yn - yo)(yn - yi) • • • (y« - y«-i) 

This relationship is important in determining the value of the 
independent variable which corresponds to a given value of the 
function. Such a procedure is known as inverse interpolation. 

7-2-5. Aitken's Method of Interpolation. 1 A practical method 
for finding a numerical value of f(x), for a given value of x, when 
several values of x and f(x) are known, is Aitken's process of 
iteration. This method is well adapted to computing machinery. 
It consists of an iteration of the familiar process of linear inter- 
polation. The basis for the following description was drawn 
largely from Aitken, 1 Feller, 6 and Milne. 16 

Eliminating all terms above the first degree in Newton's 
formula, Eq. (7-2), we are left with the expression for linear inter- 
polation, or interpolation by divided differences. Rewriting 
this expression in determinantal form, for the sake of convenience 
in computing, and substituting y x — y for Ay and X\ — x for h, 
we have 

y x — x 



*F(x) = 



yi xi — x 



(7-10) 



X\ — Xq 

The superscript (1) indicates that the function was deter- 
mined from the first iterate, and the subscript 1 indicates that 
it was formed on the difference y k — yo for k = 1. This process 
of interpolation by proportional difference is repeated to obtain 
2 1} , *i 1} , • • • , 4>n\ where 

yo Xq — X 

y k x k — x 



« 



CD = 



X k — Xq 



(7-11) 



It may be noted that, for x = x k , y = f(x) = ^ 1} . In the 

next iteration, </> (1) is substituted for y in Eq. (7-11). That is, 

the same operation performed on the set y k for k = 1, . . . , n 
is iterated on the set <j> k l) for k = 2, . . . , n, to obtain 



NUMERICAL ANALYSIS 



109 



4% 



</>i 



(i) 



X\ — x 
Xk — x 



x k — Xi 



(7-12) 



The process of iteration is continued until the computed values 
of the interpolation function 4> k n) agree to the desired number of 
places. 

In using this method to compute specific numerical values, a 
table of the following type is set up: 



Xo 


2/o 








xo — x 


Xi 


V\ 


*P 






Xi — X 


x 2 


2/2 


4" 


A,™ 
02 




Xi — X 


Xz 


2/3 


*P 


03 


<t>[ 3) 


Xz — X 



In practice it is probably most easy to form all the inter- 
polating polynomials of successively higher degrees by sys- 
tematically following the rules illustrated in Eqs. (7-11) and 
(7-12). It is evident, however, that the method can be described 
more generally by stating that each value of <f> M is obtained by 
linear interpolation applied to any two different interpolating 
polynomials </> (n-1) . 

7-2-6. Hermite's Formula for Interpolating Periodic Func- 
tions. If a function y = f(x) is periodic in the interval 2ir, it is 
appropriate to use trigonometric interpolation. A formula 
derived by Hermite, and also independently by Gauss, is appli- 
cable to this condition. Similarity to Lagrange's formula, 
Eq. (7-8), for nonequidistant values of the argument of a non- 
periodic function is quite apparent. 

sin (x — Xi) sin (x — x 2 ) ■ * • sin (x — x n ) 



y = 



+ 



sin (xo — Xi) sin (x — x 2 ) 
sin (x — Xo) sin (x — x 2 ) 



• - sin (xo — x n ) 

• • sin (x — x n ) 



2/o 



+ 



sin (xi — Xo) sin (x\ — x 2 ) 
sin (x — Xo) sin (x — Xi) 



sin (xi — x n ) 
sin (x — Xn-i) 



+ 



Vn (7-13) 



sin (x n — Xo) sin (x n ■— Xi) • * • sin (x n — #n-i) 

In common also with Lagrange's formula, Hermite's trigonometric 
interpolation formula may be transposed, by interchanging x 
and y, for inverse interpolation of periodic functions. 

7-3. Numerical Differentiation 

The process for obtaining the numerical value of a derivative 
of a given order is connected closely with the use of the interpola- 



110 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

tion formulas employing differences such as those which we have 
briefly outlined in Sec. 7-2. The obvious restrictions also apply 
in the matter of equidistant values of the argument, wherein 
the formulas of Newton, Bessel, and Stirling may be applied, and 
in values of the argument which are not equidistant, wherein 
the function is represented by formulas of Lagrange for non- 
periodic functions and of Hermite for periodic functions. 

The numerical values of derivatives are determined through 
the use of differences and interpolation formulas based on these 
differences. Taking Newton's formula for forward interpolation 
as an example: 

y = 0(u) = y + u Ay + -^j — J -A 2 y + -^ ^ J -~ A*y 

, u(u - l)(u - 2)(u - 3) A4 „, 
-r ^ ai/o -t- • • 

+ u(u-l)(u-2)^ ■ ( Un -n + l) Anyo (? _ 14) 

Since u = (x — x )/h, it follows that 

dy _ dy du _ 1 dy 
dx du dx h dx 

The values of several of the derivatives of y with respect to x 
are given below in terms of differences appearing in Newton's 
formula of interpolation. 



dy 
dx 



1/ 2u-l A2 .3u 2 -Qu + 2 AZ 
= ^ I Ay + 2 , A 2 2/o + 3j— - A*y Q 



, 4w 3 - 18w 2 + 22^-6 AA 
+ 4J A^o 



s 2/o + ' • ) (7-15) 



5w 4 - 40w 3 + 105w 2 - lOOu + 24 
5! 



d 2 ?/ 1 [ A9 , , 1W , . 12u 2 - 36u + 22 A4 
^J = }p [ A2 2/o + (i* - 1) A 3 2/o + ^ A^o 

. 20w 3 - 120w 2 + 210^ - 100 A5 . 1 /Py 1fl x 

H ^ A 5 ?/o + * ' * J (7-16) 

d 8 2/ 1 / A , , 24w - 36 A4 

" jp (^ A3 2/o H -rr - AVo 



dx* h z V * u 4! 

)u 2 - \ 

5! 



. 60^ 2 - 240w + 210 A5 \ , 7 17 , 

A 5 2/ + 1 (7-17) 



NUMERICAL ANALYSIS 111 

g - I [A<y. + (« " 2) A^„ + • • •] (7-18) 

g - i [A^o + • • •] (7-19) 

It is seen from the method of determining a numerical value for 
the derivatives of a function that this process is capable of being 
handled by machine operation where tabular values of a function 
are given. 

7-4. Numerical Integration 

This process allows the numerical evaluation of a definite 
integral. Again here, as in differentiation, the method of deter- 
mining a numerical value for a definite integral depends upon 
interpolation formulas discussed in Sec. 7-2. Since numerical 
integration is based principally on differences, it is capable of 
being performed by machine where the input data are values 
of the function and the function differences. In formulas for 
interpolation of functions with equidistant values of the argu- 
ment, the increment by which the argument is increased succes- 
sively was designated by h where x = x + hu. In differential 
form dx = h du. For n values or intervals of the argument, x 
is increased by an amount nh. If now we integrate Newton's 
formula for forward interpolation, Eq. (7-3), between the limits 
x and Xq + nh the corresponding limits for u will be and n. 
The integral of y dx may then be expressed in terms of the 
original differences of the tabulated function and n, the number 
of equidistant increments in the argument. This results in the 
following expression for this integral: 



J ydx = h \ny + -^ Aj/„ + I -j - -j I -gf 

j^( ni 3_l 2 A A3 2/o,/w 5 Zn'Un 3 A A^o 



This is a general formula from which a number of well-known 
special formulas for integration may be obtained. 19 It may also 



112 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

be written in terms of the sums of values of y n of the function; 
in this form it is known as the Newton-Cotes formula. 15 

Several so-called rules derived from the integral result, by 
allowing n to take on various values such as 1, 2, 3, and 6, are 
listed below. 

7-4-1. The Trapezoidal Rule. If we allow n to take on the 
value 1 and we neglect all differences above the first, the following 
expression is obtained for the integral of y dx : 

J ydx = h{^ + yi + yi .+ »,+ ••• + &1 (7-21) 

These are just the first terms in the Euler-Maclaurin formula. 15 

7-4-2. Simpson's One-third Rule. By putting n = 2 and 
neglecting all differences above the second, a very important 
formula known as Simpson's one-third rule is obtained: 

fxo + nh i 

I ydx = ^[yo + 4(2/i + 2/3 + * * * + 2/n-i) 

+ 2(2/2 + 2/4 + ' ' • + 2/n- 2 ) + 2/J (7-22) 

7-4-3. Simpson's Three-eighths Rule. With n = 3, and by 
neglecting all differences above the third, an expression for this 
interval known as Simpson's three-eighths rule is obtained: 



/ 



xo + nh ^ 



y dx = -g- [y + 3(2/i + 2/2 + 2/4 + 2/5 + 2/7 + 2/8 



Xo 

+ • • • + 2/n-l) + 2(2/3 + 2/6 + 2/9 + * ' * + Vn-Z) + 2/J 

(7-23) 

7-4-4. Weddle's Rule. With n = 6 and neglecting all differ- 
ences above the sixth, we obtain another important formula, 
probably the most accurate of these four, known as Weddle's 
rule: 



'xo+nh Q7 

ydx = T7 ,(y + 52/i + 2/2 + 62/3 + 2/4 + 5y b + 2y 6 



Jx * ~~ 

+ 52/7 + 2/8 + 62/9 + 2/10 + 52/n + 22/12 + 22/n-e + 52/ n -5 

+ 2/n-4 + 62/n-3 + 2/n-2 52/n-l + 2/n) (7~24) 

It is important to note here that n must be six or a multiple of six. 

Other methods for obtaining numerical values of definite 

integrals make use of the Stirling and Bessel interpolation 



NUMERICAL ANALYSIS 113 

formulas based on central differences. 19 Since these are similar 
in character to the ones which have been sketched above, it does 
not appear appropriate to include them in this chapter. Because 
of its special character, however, another method for determining 
the numerical value of an integral, due to Gauss, will be dis- 
cussed briefly here. 

7-4-5. The Method of Gauss. All the previously outlined 
methods for determining a numerical value for a definite integral 
are based on equidistant values of the argument. Where these 
values are not equidistant, however, the determination of a 
definite value of a numerical integral is best performed by using 
a formula developed by Gauss. 15 In the development of this 
formula, Gauss solved the problem of determining, for a given 
number of values of f(x), how the values of the argument should 
be distributed in the interval (a, 6) so as to give the greatest 
possible accuracy. The result obtained by Gauss indicated 
that these values should not be equidistant but should be dis- 
tributed symmetrically with respect to the mid-point of the 
interval of integration (a, b). The resulting formula of Gauss 
may be written 

1 = fa^ dx = ( b ~ 0)[fll*(«l) + R2Cf>(u 2 ) + Rs4>(u Z ) 

+ R,4>(u,) + • • • + R n 4>(u n )] (7-25) 

The values of u h u 2 , . . . , u n represent the subdivisions of the 
interval u = to u = 1. The values of x corresponding to each 
of the w's may be obtained from these relations: 

Xi = a + (b — a)ui 

x 2 = a + (b — a)u 2 (7-26) 

x n = a + {b — a)u n 

Both the quantities u and R must be determined numerically in 
order that this equation may be applied to a definite problem. 
The Gauss method gives an exact result where the order of the 
function is less than or equals 2n — 1. Numerical values of the 
u's and the #'s for n = 3, 4, 5, 6, 7 are given in Table 7-3. 

7-4-6. An Example Using Three Methods of Numerical 
Integration. As an example of the application of (1) Simpson's 
one-third rule, (2) Weddle's rule, and (3) the method of Gauss, 



114 THE BASIC ELEMENTS OF MACHINE COMPUTATION 



the integral 



/ j—. — j dx = / 3/ <te 



will be evaluated numerically. 

Application of Simpson's Rule. Both Simpson's and Weddle's 
rules demand that a tabulation of the value oif(x) = x & /(l + x 2 ) 
be made. The interval (a,b) has been divided into six equi- 
distant increments of the argument x. Table 7-4 results. 



Table 7-3. Quadrature Method of Gauss 


for Numerical Integration 




n = 3 


n = 4 


n = 5 


n = 6 


n = 7 


U4 


0.1127016654 
0.5 

0.8872983346 


0.0694318442 
0.3300094782 
0.6699905218 
0.9305681558 


0.04691007703 

0.2307653449 

0.5 

0.7692346551 

0.9530899230 


0.03376524290 

0.1693953068 

0.3806904070 

0.6193095930 

0.8306046932 

0.9662347571 


0.02544604383 
0.1292344072 
0.2970774243 
0.5 


Uf, 




0.7029225757 


Ue 






0.8707655928 


Un 








0.9745539562 


R, 

R2 

Rz 


0.2777777777 
0.4444444444 

0.2777777777 


0.1739274226 
0.3260725774 
0.3260725774 
0.1739274226 


0.1184634425 
0.2393143352 
0.2844444444 
0.2393143352 
0.1184634425 


0.0856622462 
0.1803807865 
0.2339569673 
0.2339569673 
0.1803807865 
0.0856622462 


0.06474248308 
0.1398526957 
0.1909150253 
0.2089795918 






0.1909150253 








0.1398526957 


fi\ 








0.06474248308 





















Table 7-4 






X 


X 2 


1 +x* 


X 6 


y = /(«) 


x 


1.4 


1.96 


2.96 


7.5295360 


?/o = 2.5437622 


Xi 


1.5 


2.25 


3.25 


11.390625 


yx = 3.5048077 


Xi 


1.6 


2.56 


3.56 


16.777216 


7/ 2 = 4.7127011 


Xz 


1.7 


2.89 


3.89 


24.137569 


yz = 6.2050306 


Xi 


1.8 


3.24 


4.24 


34.012224 


2/4 = 8.0217509 


X 6 


1.9 


3.61 


4.61 


47.045881 


2/5 = 10.2051803 


X* 


2.0 


4.00 


5.00 


64.000000 


2/6 = 12.8000000 



P° x 6 1 

Is = / t-t— 2 dx = ^ [2.5437622 + 4(3.5048077 
JiA 1 + z 3 

+ 6.2050306 + 10.2051803) 

+ 2(4.7127011 + 8.0217509) + 12.8000000] 

= 4.0157580 (7-27) 



NUMERICAL ANALYSIS 
Application of Weddle's Rule 



115 



f 20 x« 



dx = 0.03(2.5437622 + 5 X 3.5048077 

+ 4.7127011 + 6 X 6.2050306 

+ 8.0217509 + 5 X 10.2051803 

+ 12.8000000) 

= 4.0157507 (7-28) 

Application of the Gauss Method 

f 20 x G 
I G = / 1 , 2 dx = O.6[/2i0(wi) + R 2 <t>{u 2 ) + R 9 <l>(u t )] 

J 1.4 J- "T" X 

using n = 3. We now make use of the values of u\, u 2 , and u s 
in Table 7-4 to obtain the appropriate intervals Xi, x 2 , and x s of 
the argument. 

Xl = 1.4 + (2.0 - 1A)U! = 1.467620999 

x 2 = 1.4 + (2.0 - \A)u 2 = 1.70 

x 3 = 1.4 + (2.0 - lA)u 3 = 1.932379000 

Table 7-5 allows the determination of <f>(ui), <t>(u 2 ), and (f)(u 3 ). 

Table 7-5 



1.467620999 

1.70 

1 . 932379000 



2.153911397 

2.89 

3 . 734088599 



9.992714509 
24.13756900 
52.06595692 



In this example, cf>(u) = x*/(l -\- x 2 ); therefore 

, . 9.992714509 Qim , A1 , n 
* (ttl) = 3.153911397 = 3 - 168356130 

4>(u 2 ) = 24J f g 5 Q 690Q - 6.20503059 

,. , 52.06595692 1ftoOQnoK1 -. 
* (Ms) = 4.734088599 = 10 " 809516 

3 



Multiplying out the terms } Ri4>(ui), using the values of 12 1, R 2) 
and R 3 in Table 7-3, 



z = l 



116 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

R^m) = 0.880098923 
R 2 4>(u 2 ) = 2.757791373 
R 3 4>(u 3 ) = 3.055026432 



^ Ri<t>(ui) = 6.692916728 



3 

:) dx = (b-a)\ {Ri<t>{ui)} 
i = i 



/> 




9 dx = 0.6 X 6.692916728 

+ x 2 

= 4.0157500 (7-29) 

In order to check the accuracy of the methods worked out above, 
the integral has been evaluated to eight significant figures. 



f 20 x« J [x* X* . 

/ i ■ 2 dz = T ~~ "5" + x ~ tan 
JiA I + x 2 15 3 



2.0 
1.4 

= 4.0157501 (7-30) 



In this particular example, it is apparent that the method of 
Gauss is the most accurate, while Simpson's rule is the least 
accurate. Weddle's rule, however, gives a numerical value 
close to the true evaluation (within six digits in the eighth 
place). 

7-4-7. Evaluation of Double Integrals. There are many 
important applications in which the numerical value of a double 
integral is required and where it is not possible to perform the 
integration analytically. A definite double integral of a function 
of two independent variables may be evaluated by extending 
Simpson's rule to functions of two variables or by a repeated 
application of ordinary quadrature formulas for one variable. 
Both of these methods are straightforward in principle but may 
involve detailed numerical calculations. Graphically the pro- 
cedure amounts to evaluating the function at selected points 
on a two-dimensional grid. The intervals along the abscissa 
and the ordinate may be of the same magnitude or may differ 
greatly in magnitude, depending upon the manner in which the 
function changes with each of the independent variables. 



NUMERICAL ANALYSIS 117 

7-5. Numerical Solution of Algebraic and Transcendental 
Equations 

The methods for finding roots of a numerical equation are 
numerous. We are interested in this section in examining those 
methods for solving definite problems which may be used as 
components in a computing-machine program. Three methods 
will be discussed: the Newton-Raphson method, the method of 
false position, and the method of iteration. These methods 
apply both to algebraic and to transcendental equations. For 
two simultaneous algebraic equations, an extension of the 
Newton-Raphson method and the method of iteration will be 
discussed, and, finally, the Graeffe root-squaring method for 
solving algebraic equations. It is realized that the material 
chosen for this section does not by any means exhaust the dis- 
cussion of the entire domain of methods which could be applied 
to particular problems. 

The Newton-Raphson method, the method of false position, 
and the method of iteration for finding roots of algebraic or tran- 
scendental equations require that an approximate value of the 
root be known before these methods may be applied. A graphical 
procedure is frequently used to find an approximate value of the 
root of an equation. 14 With such an approximate value given, 
these methods essentially apply corrections to this approximate 
value, and, by this means, the successive values obtained con- 
verge to the correct root of the equation. 

7-5-1. Newton-Raphson Method. This method makes use 
of an approximate value of the desired root and the derivative 
of the function. Let us assume that the approximate value of 
the root of a function f(x) = is denoted by the quantity Xo 
and that this value must be corrected by a quantity h to give the 
exact value of the root, 

x = Xo + h 

We now expand the function f(x + h) = in Taylor's series and 
neglect all terms in h higher than the first, leaving the approxi- 
mate relation 

f(x ) + hf(x ) = 
The first approximation hi to the correction h necessary to give 



118 THE BASIC ELEMENTS OF MACHINE COMPUTATION 
the desired root is 

*--$S (7 - 3i) 

The root of the equation is then improved to a value 

si = x + h, = xo - jj^ (7-32) 

This process is continued until the value of the root is obtained 
to the desired accuracy. 

As an example of this method, consider f(x) as the polynomial 

fix) = ax 2 + bx + c = (7-33) 

The derivative of fix) is 

fix) = 2ax + b 

Using x = Xo as the first approximation and substituting in 
Eq. (7-32) yields 

A special case of Eq. (7-33) is the condition where a = 1, 
6 = 0, and c = — u, so that 

x 2 = u 
x = y/u (7-35) 

Substituting into the expression for x h the equation is obtained: 

* - 5 (* + j) (7 - 36) 

which is a useful and well-known form. 

As another example of the use of the Newton-Raphson method 
consider the transcendental equation 

f(x) =z-Ksina;-l=0 (7.37) 

(7-38) 





/'(*) = 1 - 


y% 


COS X 




Substituting in 


(7-32), 










Xo — 
X\ Xq j 


}>£ sin xo — 

— Yz COS Xq 


1 


As a first approximation take 









NUMERICAL ANALYSIS 119 



7T 

so = 2 



w _ Qr/2) - H sin Qr/2) - 1 _ L5 



Xl 2 1 - M cos (i/2) 

= _ 1.5 -^ sin (1.5 radians) - 1 = umM2 
1 — cos 1.5 radians 

For this value, /(1.498062) = 0.000616. 

7-5-2. Method of False Position. To apply this method, it is 
necessary first to find two values of the argument between which 
the function f(x) vanishes. Preferably these values should be as 
close together as possible. With these two values of the argu- 
ment and the two known values for f(x) at these two points, a 
first approximation to the root is obtained by linear interpolation 
between these two values of the function. Essentially the func- 
tion f(x) in the interval is replaced by a straight line. By 
successive applications of this method it is possible to find a value 
of the root to the degree of accuracy required. 

7-5-3. Method of Iteration. A method for the evaluation of 
f(u) which is extremely useful for many special cases is based on 
the following procedure: Given the equation 

x = }{u) (7-39) 

it is possible to choose many functions g(x,u) such that for all 
values of u 

g[f(u),u] = k 

where k is constant. For some convenient choice of g(x,u), 
consider the equation 

g(x,u) = k (7-40) 

The principal criterion of choice is convenience in writing the 
inverse (in u) of the function g(x,u) with x held fast. For 
x = xo, an approximate value of f(u), g(x ,u), may be evaluated 
and written 

g(x ,u) = k + h (7-41) 

where h is the difference between g(x,u) and g(x ,u). 

The function f(u) may now be expressed as a function of Xo 
and h by solving for u in Eq. (7-41) and evaluating f(u) : 

f(u) = F(x ,h) (7-42) 



120 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

The next approximation for x is then obtained by expanding 
Eq. (7-42) and retaining only the linear term in h. 

In this approximate solution the value for h from Eq. (7-41) 
is substituted and the result simplified. 

As an example of this method, the development of an expression 
for the square root of u will be made. For x = f(u) write 

x = \/u (7-43) 
Squaring both sides: 

x 2 = u (7-44) 

u - x 2 = (7-45) 

Equation (7-45) is an identity of the type sought, and it leads to 
a choice g(x,u) = u — x 2 and k = 0. Then, if x is an approxi- 
mate square root of u, in accordance with Eq. (7-41), write 

u - x 2 = h (7-46) 

Taking the square root, 

v* = (h + xl) H (7-47) 

and regrouping, 

uH = x ° i 1 + Fif (7_48) 

The quantity [1 + (h/x 2 )] iA may be expanded by the use of 
Taylor's series. 



M = 



For the approximation only the linear term in h is considered, so 
that f(u) = f(x ,h) becomes 



(•♦# 



u* ^ xo I 1 -f g^l (7-50) 

Substituting in Eq. (7-50) a value for h obtained from Eq. (7-46), 

U* £* xo + ^- (u - 4) (7-51) 

or 

«h«J(*+-J) (7-52) 

which agrees with Eq. (7-36) above. 



NUMERICAL ANALYSIS 121 

A different expression may be developed for y/u by use of the 
same method. From the expression 

(7-53) 

(7-54) 

(7-55) 

Hi 

Let x be the first approximation to x; then g(x ,u) = k + h 
becomes 

| = 1 + h (7-56) 

Solve for u and take the square root: 

u A = 3 (1 + h)~* (7-57) 

Again using the Taylor series: 





x = \/u 


write [to obtain g(x,u) 


= k] 




X 2 

g(x,u) = — 

to 




X 2 



\/u = 



Xq 



1-^+ 2 A) W+ ■ ■ ■ (7-58) 



And again retaining only the linear portion of this expression 
Vu = *o(l - Hh) (7-59) 

But h may be expressed in terms of u and x according to 

(7-56), and 

. V^ = Xo[l-- 2 (^-l)] (7-60) 

Collecting terms and regrouping, 

Vu^l(sxo-f) (7-61) 

Both Eqs. (7-52) and (7-61) give useful iteration schemes 
for square root of u. 

As a numerical example, it is desired to calculate -\/2 to 10 
significant figures using both Eqs. (7-52) and (7-61). 

(a) Using Eq. (7-52) : Let 

Xo = 1 
then 

x l = 1.5 



122 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

Inserting x\ as the approximate solution in Eq. (7-52) leads to 
the improved approximation 

x 2 = 1.416667 
then 

x 3 = 1.414215686 
and 

x A = 1.414213562 

which is the correct value of -\/2 to 10 places. 
(b) Using Eq. (7-61): Let 

Xo = 1 
then 

xi = 1.25 

x 2 = 1.386718750 

xs = 1.413416937 

x A = 1.414212890 

x h = 1.414213562 

which is the correct value of \/2 to 10 places. 

The need for an approximation of the value of the inverse of a 
quantity is frequently encountered. The required inverse may 
also be evaluated by using the previous method. For example, 
let 

x = - (7-62) 

v 

An analysis similar to that for -\/u may be developed. For 
g{x,v) = k, write 

vx = 1 (7-63) 

Let Xo be an approximate solution and write 

vxo = 1 + h (7-64) 

Solving for 1/v, 

1 t„ 

(7-65) 



(7-66) 
(7-67) 





1 Xo 

v 1 + h 


Using the first two terms of the h 


>inon 




v = * o(1 " 


■ h) 


Substituting for h 








- = z (2 - 

V 


Too) 



NUMERICAL ANALYSIS 123 

which is a simple easily calculable expression improving an 
approximation to the inverse of v. The application of Eq. (7-67) 
to the design of calculating machinery was first suggested by 
Aiken in 1938. It is used in the Harvard Mark II. 

The Newton-Raphson method and the method of iteration 
may be extended to find roots of several simultaneous algebraic 
and transcendental equations. The extension of both these 
methods is obvious. The scope of this book does not allow a 
detailed exposition of these methods, especially since they are 
well described elsewhere. 

7-5-4. The Root-squaring Method of Graeffe. The methods 
which have been described (Sec. 7-5-1 through 7-5-3) deal only 
with procedures for finding real roots of algebraic and trans- 
cendental equations; a further restriction on these methods is 
the requirement that an approximate value of the root of an 
equation be known before it is possible to apply any of them. 

A process has been developed by Graeffe for finding both the 
real and the complex roots of polynomials without recourse to 
initial approximate values. This method is based upon the 
transformation of a given equation into another equation whose 
roots are high powers of the roots of the given equation. The 
first step in this process is to transform a given algebraic equation 
into one whose roots are the squares of the given equation. The 
second equation is again transformed into a third equation whose 
roots are squares of the roots of the second equation and, con- 
sequently, are fourth powers of the roots of the given equation. 
This root-squaring process is applied successively until the ratio 
of any root to the next larger root is negligible in comparison with 
unity. The roots of the last transformed equation are then 
considered as separated. By this method the original equation 
is transformed into a set of simple equations from which the 
desired roots may be obtained. 

If «i, a 2 , a 3 , . . . , a n are roots of the equation f(x) = 0, 
this equation may be expressed as a polynomial. 4 

f(x) = (x — ai)(x — a 2 )(x — cl z ) • • • (x — On) (7-68) 

If this polynomial is multiplied by /( — x) where 

/(-*) = (~l) n (x + a x )(x + a 2 )(x + a,) • ■ • (x + a») (7-69) 

a new polynomial results: 



124 THE BASIC ELEMENTS OF MACHINE COMPUTATION 

Mx) =f(x)f(-x) = (-1)»(* 2 - a\){x 2 - al)(x> -«!)••• 

(x*-al) (7-70) 

This polynomial has roots which are the squares of the roots of 
Eq. (7-68). Again if Eq. (7-70) is multiplied byfi(-x), another 
polynomial f 2 (x) is obtained: 

Mx) =fi(x)fi(-x) - (^ - a\)(x± - a\){x* - «J) - • • 

(x*-a* n ) (7-71) 

with Toots which are the fourth powers of the roots a h a 2 , a 3y 
. . . , a n of Eq. (7-68). This procedure may be repeated until 
the roots are separated as defined above. 

After the rth squaring operation the result may be expressed 4 

(x 2r ) n + 6 1 (x 2 n - 1 + b 2 (x 2r ) n ~ 2 + • • • + b n = (7-72) 
If y = x 2r , 

y n + biy n-l + b2y n-2 + . . . + ^ = Q (7 _ 73) 

If the operation of squaring has been carried out so that the roots 
are properly separated, the approximate absolute values of the 
roots 2/1, 2/2, . . . , y n of Eq. (7-73) and a h a 2j . . . , a n of Eq. 
(7-68) are 



(7-74) 



The Graeffe method does not permit a determination of the signs 
of the roots a h a 2 , a 3 , . . . , a n ; the signs must be found either 
by graphing the function f(x) or by other methods such as the 
application of Descartes' law of signs. 9 

As an example of the Graeffe method a start will be made to 
find the roots of the fourth-degree polynomial 

x 4 - 10x 3 + 35x 2 - 50x + 24 (7-75) 

(which are known at the outset to be 1, 2, 3, and 4). On the 
first squaring, the eighth-degree polynomial is obtained: 



I2/1I ~ w, 


|ai| ~ \/bi 

1 1 *F* 


W-^; 


W ~ >/£ 



30x 6 + 273x 4 - 820a; 2 + 576 (7-76) 



This has roots: 



NUMERICAL ANALYSIS 

Xl = -y/30 = 5.477 
X2 = V^Wo = 3.017 



xz = V 820 Ai3 == 1-733 
Xi = V 57 %20 



125 



(7-77) 



= 0.8381 



The result of the second squaring operation leads to 

x u - 354x 12 + 26,481x 16 - 357,904x 4 + 331,776 (7-78) 
The third squaring process gives 

z 32 - 72,354x 24 + 448,510,881.t 16 - 110,523,752,704x 8 

+ 110,075,314,176 (7-79) 

A summary of the first three approximations to the roots 
«i, « 2 , a 3 , and a 4 is given in Table 7-6. 





Table 7-6 






First 
approximation 


Second 
approximation 


Third 
approximation 


Q!4 


5.477 
3.017 
1.733 
0.8381 


4.338 
2.941 
1.917 
0.9812 


4.049 
2.979 
1.991 
0.9995 



The root squaring may be continued until the desired accuracy 
in the roots is obtained. The application of the Newton-Raphson 
method with these first approximations may, in some cases, be 
found to converge faster to acceptable roots than the squaring 
process. It should be noted that the sum of the roots in the 
third approximation is 10.0185, which differs by less than 0.2 per 
cent from the value of the coefficient of the cubic term in Eq. 
(7-75). 

7-5-5. Numerical Solution of Simultaneous Equations. The 
Newton-Raphson method and the method of iteration may be 
extended to find the real roots of several simultaneous algebraic 
and transcendental equations. The extension of these methods 
to several equations is described by Scarborough, 19 by Margenau 
and Murphy, 15 and by Whittaker and Robinson. 20 

The numerical solution of simultaneous linear equations, 



12G THE BASIC ELEMENTS OF MACHINE COMPUTATION 

which are of prime importance in physics and in chemistry, may 
be performed by several methods. The method of deter- 
minants is useful 15 where the number of unknowns is less than 
four or five. Beyond this number the method becomes unduly 
arduous. Matrix methods for solving simultaneous linear equa- 
tions are described by Frazer, Duncan, and Collar, 8 by von 
Karman and Biot, 12 by Bargmann, Montgomery, and von 
Neumann, 2 and by von Neumann and Goldstine. 17 These 
methods are useful where there is a large number of unknowns. 
In quantum-mechanical problems, in which characteristic roots 
of secular determinants are required, a polynomial method and 
an iteration method are described by Margenau and Murphy. 15 

7-6. Numerical Solution of Ordinary Differential Equations 

The numerical solution of differential equations may be per- 
formed by a number of well-established methods. These 
methods are well described in the literature, 15>18 " 20 and extensive 
samples describing each of these methods are available to those 
who desire a more detailed understanding of the principal 
methods. Of the many methods which are available for general 
and for many special types of equations, four methods have been 
chosen for treatment in this chapter. These procedures are the 
methods of Picard, Adams, Runge-Kutta, and Milne. A review 
of these methods is given by Bennett, Milne and Bateman; 3 
numerical examples may be found in Scarborough 19 and in 
Whittaker and Robinson. 20 

7-6-1. Method of Picard. The method of Picard is essentially 
a process for obtaining the numerical solution of a differential 
equation by successive approximations. Let us consider initially 
a first-order differential equation in x and y. This may be written 
as follows: 

% = f(x ' y) (7 - 80) 

Transposed to integral form, this equation may be expressed as 

y = Sf(x,y) dx (7-81) 



or 



V-Jg|<te (7-82) 



If the initial values for x and y are Xo and y , 



NUMERICAL ANALYSIS 127 

V = Vo + f x J(x,y)dx (7-83) 

In order to determine successive values of y as a function of x, 
this integral equation must be solved. We start by holding y 
constant at y and actually perform the integration with respect 
to x. This leads to a value y x as a first approximation for y, or 

yi = yo+ f x X J(x,y )dx (7-84) 

Using the value of yi thus obtained, a second approximation yi 
may be found from the relation 



2/2 



Vo + fj(x,yi) dx (7-85) 



In obtaining a numerical value for this integral equation, the 
region in which the desired solution lies is divided into a number 
of increments (proceeding from the lower to the upper limit). 
When a satisfactory value for y is obtained after the first step, 
the procedure of integration is repeated again in the same manner 
starting with the values of x and y just determined as initial 
points. By this means a numerical solution for the integral 
equation, and hence for the differential equation from which it 
was derived, can be obtained. 

Another method closely associated with the method of itera- 
tion, or the method of Picard, may be performed by use of an 
approximating polynomial similar in form to that which was 
considered earlier in this chapter in the discussion of Newton's 
formula for backward interpolation. 

The application of the method of iteration can be made to 
differential equations of the second order. This method, how- 
ever, requires that the differential equation of the second or 
higher order be reduced to a system of first-order equations. 

7-6-2. Method of Adams. Although the method of Adams 
is somewhat similar in procedure to the method just described, 
it depends for its usefulness on the ease with which derivatives 
of the function can be obtained. If y = f(x) is the desired 
integral relationship appropriate to a differential equation for 
which a numerical solution is desired, the method of Adams 
proceeds by expanding this function in a Taylor series about Xq: 



128 THE BASIC ELEMENTS OF MACHINE COMPUTATION 
f(x) = /(*„) + (X- Xo)f(x„) +(x- x„y f ^ff- + • ■ ■ 



+ {x - Xo) J-^ 
In terms of y, this expansion is equivalent to 



(7-86) 



y = 2/0 + 2/o(* - xo) + y ( x - x «) 2 + • • • + ^r ^ ~ *°) n 

(7-87) 

In order now to find approximate values for y, namely, y h y 2 , 
2/3, 2/4, etc., we next determine the derivatives y Q , y", y' Q n from the 
given differential equation for the initial values x = x and 
y = y ; we take the increments in x, which are h, 2h, etc., 
successively in the last equation. 

The success of the method of Adams depends on how easily the 
required derivatives may be determined and how quickly the 
Taylor series converges. It is pointed out, for example, by 
Scarborough, 19 that this method is not applicable to starting 
the computation of a ballistics problem because of the difficulty 
in calculating successive derivatives of the air-resistance term. 

7-6-3. Method of Runge-Kutta. This method differs from 
either of the methods previously described in that the increments 
of the function are calculated only once by means of definite 
relationships which will be given below. In this sense, the 
numerical solution of a differential equation by this method is in a 
series of approximations for any single increment. Consider a 
differential equation of the first order in this form: 

% = f(x,y) (7-88) 

and let the quantity h be taken as the increment by which the 
independent variable x is attained. It is required to find the 
increment by which y changes from the initial value y . By the 
Runge-Kutta method, the formulas which are applied are given 
below: 

ki = f(x ,yo)h 

k 2 = / ( xo + ^ 2/o + y ) h 

h ^/Uo+^yo+^Jh (7-89) 

&4 = f(x + h, 2/0 + £3)^ 
A?/ - M(fci + 2fc 2 + 2A; 3 + k A ) 



NUMERICAL ANALYSIS 129 

The last equation determines the increment in y as a function of 
k\, &2, k 3 , and & 4 . By this procedure then, values for x\ and yi 
determined by the initial conditions x and y , and the denning 
equations above, are 

xi = x + h; 2/1 = 2/0 + Ay '7-90) 

Successive values y 2 , y$, y±, etc., are determined in the same way 
and with the use of the same formulas, with x x replacing x and y^ 
replacing y . It is interesting to note that if the derivative is a 
function only of x, this method reduces to Simpson's rule. 

The method of Runge-Kutta may be extended to simultaneous 
first-order equations or to higher order equations reduced to a set 
of first-order differential equations by introducing auxiliary 
variables for transformation. The principal advantages in this 
method over the method of Picard or the method of Adams lie 
in the fact that there is required no process of approximation of 
lengthy iteration by which each increment in the dependent 
variable is obtained. 

7-6-4. Milne's Methods. An interesting method has recently 
been developed by Milne for the numerical solution of differential 
equations. This method allows values of y to be calculated after 
the first four values of y and y' have been determined by other 
means, perhaps by one of the several methods described above. 
With these values, the Milne method can then be applied to 
determine further successive values of y by using an integral 
form of Newton's formula for forward interpolation, Eq. (7-3), 
expressed in terms of y' (see Sec. 7-2-1). For example, the next 
value of y may be approximated by the use of this formula 

Ah 
iff = 2/n-4 + y (22/U - 2/L2 + 2^_ 3 ) (7-91) 

This equation is obtained by integrating from rr n _ 4 to x n (an 
interval of Ah). The value of y™ thus obtained is then sub- 
stituted into the original equation 

tx = f(x ' v) (7_92) 

so that the value of y' n may be determined. With this value of y n 
another value of y n , namely, y^\ is obtained through the use of a 
second relation obtained from Newton's formula, Eq. (7-3), by 



130 THE BASIC ELEMENTS OF MACHINE COMPUTATION 
integrating from x n -i to x n (an interval of 2h), 

y™ = y n _ 2 + | (y' n + \y' n _ x + y' n _ 2 ) (7-93) 

This allows a second value for y n to be determined and to be com- 
pared with y„\ If these values agree to the accuracy required, 
y™ is considered to be correct. It is then substituted in the 
original equation to get the correct value for y' n . By this method, 
succeeding values of y may be determined. 

REFERENCES 

1. Aitken, A. C, On Interpolation by Iteration of Proportional Parts, 
without the Use of Differences, Proceedings of the Edinburgh Mathemati- 
cal Society, Series 2, Vol. Ill, Part I, pp. 56-76 (March, 1932). 

2. Bargmann, V., D. Montgomery, and J. von Neumann, Solution of 
Linear Systems of High Order. Report prepared under Contract Nord- 
9596 between the Bureau of Ordnance, Navy Department, and the 
Institute for Advanced Study (Oct. 25, 1946). 

3. Bennett, A. A., W. E. Milne, and H. Bateman, Numerical Integration 
of Differential Equations, Bulletin 92 (The National Research Council 
of The National Academy of Sciences, Washington, D.C., 1933). 

4. Bocher, M., Introduction to Higher Algebra (The Macmillan Company, 
New York, 1907), pp. 17-18. 

5. Crumpler, T. B., and J. H. Yoe, Chemical Computations and Errors 
(John Wiley & Sons, Inc., New York, 1940). 

6. Feller, W., On Ailken's Method of Interpolation, Quarterly of Applied 
Mathematics, Vol. I, No. I, pp. 86-87 (April, 1943). 

7. Ford, L. R., Differential Equations (McGraw-Hill Book Company, Inc., 
New York, 1933), pp. 113-146, 147-162. 

8. Frazer, R. A., W. J. Duncan, and A. R. Collar, Elementary Matrices and 
Some Applications to Dynamics and Differential Equations (Cambridge 
University Press, New York, 1946), pp. 148-155. 

9. Hall, H. S., and S. R. Knight, Higher Algebra (Macmillan & Co., Ltd., 
London, 1946), pp. 459-460. 

10. Ince, E. L., Ordinary Differential Equations (Longmans, Roberts and 
Green, London; reprinted by Dover Publications, New York, 1944), 
pp. 540-547. 

11. Jeffreys, H., and B. S. Jeffreys, Methods of Mathematical Physics (Cam- 
bridge University Press, New York, 1946), pp. 237-284. 

12. von K&rman, T., and M. A. Biot, Mathematical Methods in Engineering 
(McGraw-Hill Book Company, Inc., New York, 1940), pp. 196-204. 

13. Laderman, J., and M. Abramowitz, Application of Machines to Differ- 
encing of Tables, Journal of the American Statistical Association, Vol. 
41, pp. 233-237 (June, 1946). 

14. Levy, H., and E. A. Baggott, Numerical Studies in Differential Equa- 
tions, Vol. I (C. A. Watts & Co., Ltd., London, 1934). 



NUMERICAL ANALYSIS 131 

15. Margenau, H., and G. M. Murphy, The Mathematics of Physics and 
Chemistry (D. Van Nostrand Company, Inc., New York, 1943). 

16. Milne, W. E., Numerical Calculus (Princeton University Press, Prince- 
ton, N.J., 1949). 

17. von Neumann, J., and H. H. Goldstine, Numerical Inverting of Matrices 
of High Order, Bulletin of the American Mathematical Society, Vol. 53, 
No. 11, pp. 1021-1099 (November, 1947). 

18. Runge, C, and Konig, Numerisches Rechnen (Springer- Verlag, Berlin, 
1924). 

19. Scarborough, J. B., Numerical Mathematical Analysis (Johns Hopkins 
Press, Baltimore, 1930). 

20. Whittaker, E. T., and G. Robinson, The Calculus of Observations 4th ed. 
(Blackie & Son, Ltd., Glasgow, 1944). 



Part II 
COMPUTING SYSTEMS 



The following five chapters will deal with present and proposed 
equipment for performing mathematical computations. Chapter 
8 describes the basic functions which the desk calculating machine 
can perform. Chapter 9 outlines the available types of punch- 
card equipment and indicates how these may be used to perform 
the four basic arithmetic operations. A general description of 
several large-scale digital equipments is given in Chap. 10. 
Some general remarks on analog computing machinery are made 
in Chap. 11. A possible form which a digital computer might 
take, based on the requirement for solving a certain class of 
mathematical problems, is given in Chap. 12. 

The purpose of these five chapters is to define, by example, 
the field of application of computing systems which are composed 
of the basic arithmetic and physical components to which this 
volume is primarily devoted. 



CHAPTER 8 
DESK CALCULATORS 

8-1. Introduction 

The commercial desk computing machine is a compact instru- 
ment capable of performing the basic arithmetic operations 
quickly and simply. These devices are not intended to compete 
with electronic machines of large capacity and high speed which 
are now in existence or in design. The desk calculator makes 
available to the individual computer a simple machine capable 
of straightforward calculations. With this machine, the arith- 
metic operations of addition, subtraction, multiplication, and 
division can be performed. Although desk calculators were 
developed primarily for use in the commercial and accounting 
fields, they may be used efficiently without modification for many 
scientific computations. While it is not sensible to discard as 
hopeless those computations which could be carried out only on 
large-scale digital computing equipment, it is equally unwise (as 
Comrie 1 points out) to procure expensive custom-built equip- 
ment when one of the simpler mass-produced computing devices 
will carry out the required computations more economically. 
For this reason it seems desirable to outline here the capabilities 
of standard electrically driven desk calculating machines. 

For some applications, particularly those which do not require 
the handling of a large number of digits, key-actuated non- 
automatic machines are satisfactory. A discussion of machines 
of this type is beyond the scope of this chapter, which is confined 
to descriptions of the basic functions of electrically driven desk 
calculators. 

In the United States, five companies manufacture electrically 
driven desk calculating machines: Friden Calculating Machine 
Co., Inc., Burroughs Adding Machine Co., Marchant Calculating 
Machine Co., Monroe Calculating Machine Co., Inc., and 
Remington Rand, Inc. These machines are all generally 
similar with keyboard setting, electric drive, automatic multi- 

135 



136 COMPUTING SYSTEMS 

plication and division, and a few additional special characteristics. 
They differ from each other in the ease of carrying out particular 
operations, in available storage and transfer facilities, but not in 
the type of results obtainable or in the principle of operation. 

8-2. General Description of the Machines 

The desk calculator is designed with a keyboard of several 
columns, each column containing a key for each of the digits from 
through 9. Near the keyboard are mounted various operating 
bars, clearing keys, and reversing levers; the number of these and 
the type depend upon the complexity of the machine. (Some 
machines have supplementary keyboards for the entry of mul- 
tipliers or other numbers.) Above this keyboard is a movable 
carriage with two registers; one is known as the multiplier register 
and the other as the product register. The product register 
usually has a digit capacity twice that of the multiplier register. 
Various supplementary markers and pointers are available for 
convenience in indicating decimal points, the position of the 
carriage relative to the keyboard, etc. The carriage contains 
knobs for clearing each register, and sometimes provision for 
setting numbers into the product register manually. 

8-3. Operation of the Machines 

The operation of desk calculators to carry out the elementary 
arithmetic processes is briefly as follows: 

8-3-1. Addition and Subtraction. Assume first that the 
augend or minuend is already in the product register. The 
movable carriage is placed in a position so that the digits in this 
number are directly above the corresponding digits on the key- 
board, and the addend or subtrahend is set digit by digit into 
various columns of the keyboard. Depression of the plus bar 
for one cycle of the machine adds the number entered on the 
keyboard to the number in the product register, leaving the sum 
in the product register. Depression of the minus bar for one 
cycle of the machine subtracts the number in the keyboard from 
the number in the product register and leaves the difference in 
the product register; negative numbers are represented by 
complements. For example, if all the dials are cleared and a 
given number is set into the keyboard and subtraction performed, 
the machine represents this negative number by indicating, in the 



DESK CALCULATORS 137 

product register, another number to which the given number must 
be added to give a certain power of 10. In a 20-digit product 
register this number is usually 10 20 , which exceeds the capacity of 
the machine by one digit. In some machines capable of holding 
20 digits in the product register, this number may be 10 14 . This 
system of representing negative numbers by their complements 
is commonly used for computations with logarithms in order 
to keep all mantissas positive; it serves the same purpose on large- 
scale computing machines. Several desk calculators include a 
button near the keyboard marked add; if this button is depressed, 
the keyboard is cleared after one machine cycle, and careless 
prolonged depression of the operating bars will not produce 
incorrect answers. 

8-3-2. Multiplication. To multiply, the multiplicand is set 
in the keyboard, and the product and multiplier registers are 
cleared. The machine multiplies by successive additions. With 
the carriage in its extreme left-hand position and with the add 
button released, the plus bar is depressed for a period of time 
sufficient to permit the extreme right digit of the multiplier to 
build up in the multiplier register. In this operation the digit 
of the multiplier register immediately above the extreme right 
column of the keyboard increases by one for each cycle the 
machine operates. Thus, at the time of this first step, in which 
the right digit of the multiplier is entered into the multiplier 
register, the multiplicand is added into the product register by a 
number of times equal to the right digit of the multiplier. Next, 
the carriage is shifted one digit to the right, and the process is 
repeated until the 10's digit of the multiplier appears in the 
proper place in the multiplier register. In this way, iteration of 
the add-and-shift technique yields the final product in the 
product register. 

Many keyboard machines perform automatic multiplication. 
In this process the multiplier is set into an auxiliary keyboard, 
and the multiplicand is set into the upper keyboard. The 
machine will automatically carry out the process described above. 

8-3-3. Division. Normally, division is performed by a process 
of successive subtractions similar to the multiplication process 
described above. The dividend is set in the product register well 
to the left. The divisor is set into the keyboard underneath the 
dividend. The multiplier register is cleared. A reversing lever 



138 COMPUTING SYSTEMS 

near the keyboard is thrown so that the number in the multiplier 
register increases when the minus bar is depressed and decreases 
when the plus bar is depressed. With this arrangement the 
minus bar is depressed for a time long enough to introduce a 
series of 9's at the extreme left end of the product register. 
This indicates that the partial remainder has become negative 
and that the number in the multiplier register is greater than the 
quotient sought. The plus bar is depressed for one cycle, making 
this partial remainder positive again, and the carriage is shifted 
one digit to the left. With the carriage in this position the 
process is repeated. This iteration is continued until the carriage 
is in its extreme left position. At this time the quotient is in 
the multiplier register, and the remainder is in the product 
register. This is the process described in terms of elementary 
computing operations in Chap. 6. 

This process is built in for automatic operation in many desk 
calculators. Provision for automatic division is far more com- 
mon than automatic multiplication in desk calculators. 

8-4. Capacity, Speed, and Capabilities 

In describing desk machines the capacity is usually given by 
a series of three numbers. The first of these numbers indicates 
the number of columns in the keyboard; the second indicates the 
number of digits in the multiplier register; the third indicates the 
number of digits in the product register. Thus, a machine 
described as 10 X 11 X 21 is one with 10 columns in the key- 
board, 11 digits in the multiplier register, and 21 digits in the 
product register. This is a fairly common size of machine. 
The speed of calculation is dependent upon the cyclic speed of 
the machine and upon special features built into machines. 
Most machines operate at a rate of 500 to 650 cycles per minute, 
although some models operate at a rate of 1,350 cycles per minute. 
Some time saving may be obtained if the machine starts an 
automatic multiplication while other digits are entered. Time 
may also be saved by automatic short cutting in multiplication; 
this process shortens the number of steps required in some 
multiplications by running the machine backward to get high 
digits on the multiplier register and later correcting left-hand 
digits for unwanted carry-over. Thus to multiply by 89, the 



DESK CALCULATORS 



139 



machine might first multiply by — 1, shift and multiply by — 10, 
shift again and multiply by 100 to clear the unwanted 9's. 

Since the speeds of various computing machines and their 
separate components have been mentioned elsewhere in this book, 
it seems advisable here to note briefly the performance of desk 
calculators in this respect. In the succeeding section of this 
chapter the addition speeds of several desk calculators are given. 
For two of the machines, the Friden and the Monroe specifically, 
the addition speed is of the order of 500 to 650 counts per minute ; 
for another machine, the Marchant, the addition speed is about 
1,350 counts, over twice the speed of either of the first two 
machines. 

In the operations of multiplication and division, which are 
performed on all these machines by an iterative process of suc- 
cessive addition (or subtraction) and column shift, there is no 
marked difference in the total times of all three machine opera- 
tions. To indicate this, there are listed in Table 8-1 a few repre- 
sentative operations involving 10-digit multiplications and 
divisions with the approximate performance times in each 
machine. 

Table 8-1 





Friden 


Marchant 


Monroe 


Addition speed 


500 digits 


1350 digits 


650 digits 




per minute 


per minute 


per minute 


Multiplication : 








9,999,999,999 


12 seconds 


10 seconds 


11 seconds 


X 9,999,999,999 








Division : 








9,999,999,999 
3.141592654 


9 seconds 


8.5 seconds 


10 seconds 


9,999,999,999 

2.8182846 


9 seconds 


8 seconds 


9 . 5 seconds 



It has been estimated, as a conservative figure, that in an 
8-hour day one good operator can produce 400 10-by-10-decimal- 
digit multiplications or divisions. This includes time for 
recording the answer of each operation and for making some 
elementary accuracy checks. One thousand additions may be 
performed in the course of an 8-hour day by the one calculator. 
Considerable saving of time can be effected if a program can be 



140 COMPUTING SYSTEMS 

devised in which the numbers in the registers of the machine 
are reused without transcription. For example, in using Aitken's 
interpolation method described in Sec. 7-2-5, the computed value 
of the determinant appears in the main dials and is divided by 
the appropriate factor without transcription. 

8-5. Special Features 

The manufacturers of calculating equipment have built into 
some of their machines several special features designed to 
facilitate particular types of computation. A description of a 
few of these features is given here. It should be emphasized 
that an increase in the number of automatic features on the 
machine usually represents an increase in the service require- 
ments of the device. The most economical machine which will 
carry out an expected computation is generally the best choice for 
that computation. Special features mentioned by the various 
companies are as follows: 

8-5-1. Burroughs. The Burroughs Adding Machine Com- 
pany lists two Electric Duplex Calculators which are adaptations 
of the adding machine. These calculators have no provision for 
automatic multiplication or division; these operations must be 
performed by manual positioning in steps across the keyboard to 
simulate the action of a movable carriage. The outstanding 
characteristic of these calculators is the provision for using two 
accumulating registers, one for the results of detailed calculations, 
including the four arithmetic operations, and the other for 
accumulating the results occurring stepwise on the first register. 
The results from the first register may be added to or subtracted 
from those in the total or second register; this transfer resets the 
first register for the next operation. 

8-5-2. Friden. The Friden has a 10-key setting for the mul- 
tiplier in automatic multiplication. The machine, however, does 
not have complete 10's transmission in the product register; 
essentially, the Friden takes complements on 10 14 and not 
on 10 20 . 

The Friden machine is available with both the split and the 
normal dial clearance. With this adaptation, the upper dial 
may be split into a right and left section at the seventh, eighth 
or ninth digit. The split feature enables the operator to clear 



DESK CALCULATORS 141 

the left upper dial and to accumulate in the right upper dial or to 
accumulate in the left and clear the right upper dial. 

A Multiplier Nonentry Control available only on Friden 
Models S and ST is an important advantage for calculations such 
as ab/c. With a special counter control key, the individual or 
the accumulated multipliers are prevented from entering the 
lower dial. The product ab, entered in the upper dial, may then 
be divided by c with the quotient entering the lower dial. In 
this operation only the results of division are registered and 
accumulated on the lower dial. 

The normal operation of the Friden is approximately 500 
counts per minute. 

8-5-3. Marchant. The newer models of the Marchant operate 
at a counting speed of 1,350 counts per minute. In the other 
types of machines, both multiplicand and multiplier must be 
set before multiplication begins. In the Marchant the multi- 
plicand is set as usual. The multiplier is set into an auxiliary 
row of 10 keys. As each key is pressed, its digit is absorbed into 
the multiplier and the carriage moved to the next position. 
While this machine is running on one digit, the next may be set. 
Since the speed of the machine is 1,350 counts per minute, the 
product is always completed within 3^2 second after setting the 
last digit. This feature enables an operator to carry out more 
multiplications per hour because he does not have to spend as 
much time waiting for the machine. For the same reason, 
however, some operators find the machine more tiresome to 
use for extended periods of time. 

The high speed of the Marchant is possible because the numeral 
wheels are operated by continuously driven gears. These gears 
are selected by the keyboard setting. 

„ The Marchant has also the split-dial feature mentioned above. 
Some special results have been obtained by mounting two Mar- 
chants together, the result being called a twin Marchant. Both 
machines can be driven separately, can be turned together in the 
same direction, or can be turned in opposite directions. An 
expression of the form ab/c can be evaluated in one operation. 
Twin machines have been applied to military rectangular- 
coordinate survey work and to the interpolation of pairs of 
double-entry tables. 



142 COMPUTING SYSTEMS 

8-5-4. Monroe. Some models of the Monroe Calculator 
have an automatic feature which makes it possible to form 
2a 2 with a single setting of the values a. Some older models 
of the Monroe incorporate automatic short cutting. 

Another useful feature of the Monroe is the ability to call a 
stored constant multiplier at will instead of setting it when 
required. This characteristic is useful for computing expressions 
such as Kx + y where x and y are variables and K is a constant. 
The storage of this constant in no way limits the use of the key- 
board for other operations. By use of a special short cut a con- 
stant dividend may also be stored. 

The product abc can be taken very easily on some Monroe 
machines by using a transfer switch which transfers a number 
from the product register to the keyboard. This transfer facility 
is also useful in computing such expressions as ab/c. 

Models are available with split dials. This feature enables 
the operator to operate on either half of the dial or on both halves 
and to clear either half as desired. 

In normal operation the Monroe Calculators add digits at the 
rate of about 650 per minute. 

8-5-5. Remington Rand. Remington Rand manufactures 
two printing calculators, one with a 13-column list, or keyboard, 
and 14-column total, and one with a 10-column list and 10-column 
total. The principal advantage in these calculators is their 
provision for printing all steps in the basic arithmetic operations. 
The unit operation time, including that for introducing the 
problem and for printing, compares favorably with other com- 
mercial calculating machines which do not have the printing 
feature. The 13-column model provides a special feature for 
short-cut multiplication by which it is possible to use comple- 
ments on 9's for digits of the multiplier greater than five. The 
provision for printing all factors in an arithmetic operation and 
the result of the operation provides a valuable means by which a 
check on the individual steps in any calculation, or series of 
calculations, may be obtained for a permanent record. 

8-6. Summary 

The desk calculator is a machine which will produce products, 
quotients, sums, and differences of numbers quickly and con- 
veniently. It will handle augends, minuends, products, divi- 



DESK CALCULATORS 143 

dends, sums, and differences to a capacity of about 20 digits, 
and multipliers, multiplicands, addends, subtrahends, divisors, 
and quotients up to a capacity of 10 digits. Such a machine in 
the hands of a competent operator can produce 400 full-length 
products or 1,000 sums during an 8-hour working day. The 
desk calculator is an economical method for performing com- 
putations of this order of magnitude. Its speed is low enough to 
permit continuous exercise of judgment by the operator. 

Information concerning desk calculators is readily available 
from the manufacturers. In addition, the following publica- 
tions 1-6 are noted as pertaining to the general characteristics 
and applications of these machines. 

REFERENCES 

1. Comrie, L. J., The Application of Commercial Calculating Machines to 
Scientific Computing, Mathematical Tables and Other Aids to Computa- 
tion, Vol. II, No. 16, pp. 149-159 (October, 1946). 

2. Hutchinson's Technical and Scientific Encyclopedia, Calculating 
Machines, Vol. I (The Macmillan Company, New York, 1936), pp. 315- 
319. 

3. Marchant Calculating Machine Company, Marchant Methods, Index of 
Marchant Methods and Tables Issued to Nov. 1946 Relating to Basic and 
Statistical Mathematics, MM-229 Mathematics (Marchant Calculating 
Machine Company, Oakland, Calif., 1946). 

The following titles of methods applicable to the Marchant Calculator 
have been abstracted from this reference to indicate the range of useful- 
ness of desk calculators: 

a. Roots and Powers: 

MM-88 Approximation Method for Extraction of Any Root 

MM-95 Square Root by Successive Odd Numbers 

MM-222 Fifth Root to 5 Places 

MM-302 Square Root by Iteration from Slide-Rule Approximation 

Table 68 Cube Root to 5 Places, with Extension to 9 Places 

b. Algebraic Equations: 

MM-182 A Short Method of Evaluating Determinants and Solving 
Systems of Linear Equations with Real or Complex 
Coefficients, by Prescott D. Crout, Ph.D. 

MM-183 Notes on Marchant Calculator Application to the Crout 
Method of Solving Simultaneous Equations (see MM-182) 

MM-225 Birge-Vieta Method of Finding a Real Root of Rational 
Integral Function 

MM-226 Setting up an Approximating Polynomial of Degree "n" 
from Equidistant Tabulated Values of a Function 



144 COMPUTING SYSTEMS 

MM-235 The Nogrady Method of Solving Cubic Equations 
MM-289 Evaluation of a Polynomial 

c. Interpolation: 

MM-64 Direct Curvilinear Interpolation with Lagrange Coeffi- 
cients 
MM-152 Direct Curvilinear Interpolation, Assuming Constant 

Second Differences 
MM-189 Direct Interpolation — Straight-Line and Curvilinear 
MM-209 Curvilinear Interpolation with Unequal Intervals of 

Argument. ("Divided difference" method using 

Lagrange-Newton formula.) 
MM-220 Inverse Curvilinear Interpolation — Short-Cut Method 

Includes Effect of 2nd Differences Only. (Using BessePs 

central-difference formula). 
MM-221 Inverse Curvilinear Interpolation and Finding of Roots 

of Tabulated Function. (The Comrie "two-calculator" 

method, using Bessel's central-difference formula.) 
MM-228 7-Place Lagrange 5-Point Interpolation Coefficients for 

Values of P from to 2, with Argument to 0.001 
MM-317 The A. C. Aitken Method of Curvilinear Interpolation. 

(For equal or unequal intervals of the argument.) 

d. Numerical Integration and Solution of Differential Equations: 
MM-167 Moment of Inertia of Sections Composed of Rectangular 

Areas. (A systematic work sheet for computing the 
constants of structural shapes.) 

MM-215 Area below Curve for Fractional Portion of Distance 
between Equidistant Ordinates. 

MM-216 Milne Method of Integration of Ordinary Differential 
Equations. (Complete explanation and systematic work 
sheets for this popular method. Includes much hereto- 
fore unpublished information.) 

e. Statistical Method, and Least Squares: 

MM-45 Pearson Correlation Coefficient. (With formula espe- 
cially adapted to calculator computation.) 

MM-119 Linear "Least Squares" Line of Regression and Coeffi- 
cient of Regression 

MM-242 Summations of X, XY, and IF 2 . (A special simplified 
method.) 

MM-245 Hansen- Ahlberg Method of Extending Parabolic Curves. 
(A method based upon constant second differences for 
rapidly extrapolating any second-degree function.) 

MM-314 Summations for Linear Multiple Correlation 

4. Meyer, zur Capellen, W., Mathematische Instrumente (Becker and Erler 
Kom.-Ges., Leipzig, 1944; lithoprinted by Edwards Bros., Inc., Ann 
Arbor, Mich., 1947), pp. 53-131. 



DESK CALCULATORS 145 

5. Murray, F. J., The Theory of Mathematical Machines (King's Crown Press, 
New York, 1947). 

6. Office Machines Research, Inc., American Office Machines Research Service 
(The Office Machines Research, Inc., New York, 1938-1939). Three 
binders: Adding and Subtracting, Index 3.0, Binder II. Listing Adding 
Machines, Index 3.2; Single Register Listing Adding Machines, Index 
3.21; Multiplying and Dividing, Index 4.0; Calculating Machines, Index 
4.3; Binder III. (This service is no longer available in the United States. 
Further information on the above subjects may be obtained from the 
International Office Machines Research Corp., Amsterdam, Holland.) 



CHAPTER 9 
PUNCHED-CARD COMPUTING SYSTEMS 

9-1. Introduction 

These systems of computation make use of commercially avail- 
able punched-card processing equipment. Business operations 
with such equipment are well known and have been developed 
extensively in Great Britain and in the United States over the 
past 50 years. The punched card now in use is an adaptation 
of the card originally conceived by Hollerith, who, as early as 
1889, described tabulating machinery for the punched-card 
processing of statistical data. 

Since the punched-card equipment and its use on scientific 
problems have been described adequately elsewhere, it is the pur- 
pose of this chapter to present merely an outline of a calculating 
system using punched cards. Although punched-card machinery 
was primarily developed for business accounting applications, it 
has been successfully utilized, without major alteration, in 
extensive computations of a scientific and statistical nature. 
For scientific calculations, a group of four or five standard busi- 
ness equipments, performing different functions, are usually 
operated as a computing unit. 

There are several advantages to be gained through the use of 
punched-card equipment in scientific work. In the first place, 
for many extensive calculations to which it may be applied, this 
type of equipment provides a speed in excess of that attainable 
with desk calculating machines. Second, greatly increased 
accuracy may be attained through the automatic operation with 
punched-card equipment, even in those computations where the 
speed is comparable to that of a desk calculator. 

The increased accuracy attainable with punched-card equip- 
ment stems principally from three sources. First, there is the 
possibility of automatic transcription of results into tabular 
form. Where a lengthy transcription is required, it is possible 
with punched-card equipment to print tables which can be 

146 



PUNCHED-CARD COMPUTING SYSTEMS 147 

reproduced by a photo-offset process without any intermediate 
manual duplication. This operation greatly diminishes the 
possibility of making a printing error which may go undetected 
in normal proofreading. Such a system of duplication is now 
standard with many punched-card installations and has been 
used extensively in the computation and tabulation of mathe- 
matical functions. 

A second source of increased accuracy in computations made by 
punched-card machinery lies in the fact that its automatic opera- 
tion does not weary of tedious iteration. The lessened depend- 
ence on the alertness of the operator is a considerable factor in 
increasing the reliability of the results even though the computa- 
tions on punched-card machinery may be performed no faster 
than on a desk calculator. 

A final source of increased accuracy by punched-card devices 
lies in the possibility of including automatic checking features 
which are far more reliable than a human proofreader. Use of 
these devices enables vast reduction in the probability of machine 
error, and in many cases checking equipment may be used in the 
same way to reduce the probability of undetected operator error. 

In the United States there are two principal manufacturers 
of punched-card equipment: International Business Machines 
Corporation and Remington Rand, Incorporated. Descriptions 
of the equipment now available and the equipments planned for 
the immediate future are available from these companies; 
detailed suggestions and instructions concerning the efficient 
use of machines built by these companies may be obtained from 
them on request. For scientific computations, at least, installa- 
tions of International Business Machines Corporation equip- 
ment seem more numerous than those of Remington Rand equip- 
ment. The description of machines in this chapter will therefore 
concern itself largely with IBM equipment. Comparable 
punched-card equipment of the Remington Rand type will 
also be described where it appears that it may be applied to 
scientific calculations. It should be noted that it is not the 
intention of this chapter to provide a sufficient description of 
any commercial piece of equipment to suggest its acquisition by 
any activity. 

The description of punched-card equipment given here is 
included for several reasons. One reason is the desire to assist 



148 COMPUTING SYSTEMS 

projects, faced with computing problems, in avoiding undue 
investment in needlessly complicated computing equipment 
because of the lack of information concerning the potentialities 
of punched-card equipment. Second, the manner in which 
punched-card machinery permits increasing automatic computa- 
tion is useful in connection with the study of more extensive 
computing equipment. An examination of the methods of 
computation already developed for punched-card equipment 
might be instructive to many activities, even though their com- 
puting problems are too extensive for adequate treatment by 
these machines. Finally, it is of interest to note the develop- 
ment of techniques for the utilization of this automatic equip- 
ment in computations. The ingenuity which has been exercised 
by the various workers in formulating methods with which their 
problems can be treated on this machinery, and the publication 
of these methods in the ordinary scientific literature, may indicate 
that true radical advances in the computing art might follow 
the acquisition of equipment which is somewhat less powerful 
than seems necessary at first glance. 

9-2. The Punched Card 

The primary unit for the automatic operation of these cal- 
culating equipments is the punched card. The IBM and the 
Remington Rand machines operate on punched cards of the same 
size, 3.250 inches by 7.375 inches, and the same thickness, 0.0067 
inch. Generally one corner is cut to facilitate card handling, 
matching, filing, etc. The IBM and Remington Rand cards do 
not store the same quantity of numerical and/or alphabetical 
information; there is also a different distribution of fields and 
zones on each type of card. It is important to point out here, 
however, that both IBM and Remington Rand manufacture 
devices for transcribing data from one type of card to the other. 

9-2-1. The IBM Card. The IBM card has 80 vertical col- 
umns each with 12 punching positions. The 10 lower positions 
are assigned to the digits to 9; the top two positions designated 
x and y, or more commonly referred to as 11 and 12 holes, are used 
for special coding, such as the indication of negative numbers or, 
in combination with one of the digits 1 to 9, for alphabetical 
representation. For use in specific calculations, a card may be 
divided into sections, or groups of columns, known as fields; 



PUNCHED-CARD COMPUTING SYSTEMS 



149 



this defines that portion of the card in which information of a 
certain kind will always appear. In one type of computation, 
for example, the 80 columns were divided into eight groups of 10; 
each of these groups in turn was subdivided into one column and 
three groups of three columns. 

The 80 columns are usually punched and read independently, 
each punched column denoting one digit, one letter of the alpha- 
bet, or a special character such as the algebraic sign of a number. 
For numerical representation, one perforation is required for each 
digit. For alphabetical representation, two perforations in a 
single column are used for each letter; one of these is a zone 
punch (0, 11, or 12) while the other perforation is made in the 
position identifying one of the digits 1 to 9. Table 9-1 illustrates 
the coding used to represent the alphabet on the IBM card. 

Table 9-1 



\ Zone 








\ 


11 


12 





Digit\ 








1 


A 


J 




2 


B 


K 


s 


3 


C 


L 


T 


4 


D 


M 


U 


5 


E 


N 


V 


6 


F 





W 


7 


G 


P 


X 


8 


H 


Q 


Y 


9 


I 


R 


Z 



Note that the (0,1) position is not used for the alphabet and is 
therefore available for special coding. 

All IBM computing equipment is capable of operating on data 
in punched cards. Some machines, however, can also handle 
data represented as graphite deposits on cards. These marked 
cards are of different capacity and design but are not radically 
different functionally. Both sides of a marked card are available 
for storage with 27 columns of information per side. 

The punched card is the fundamental unit of any IBM auto- 
matic computation system. Numbers computed in the course 
of a computation are punched on a card for further use, for trans- 
fer from one unit to another, or for storage. The cost of these 



150 COMPUTING SYSTEMS 

cards is about $1 per 1,000; the cards used as interim storage 
or as a transfer medium in the course of a computation are not 
generally used again. 

The face of the card may be printed to facilitate identification 
by visual reading. Some punch equipments print on the card 
the information which has been punched there; the printed 
characters usually lie above the column in which the information 
is punched. Additional equipment is also available for reading 
information punched on a card by other equipment and for 
printing some of this information at the edge of the card. Thus, 
the cards may serve as stencils with no further use of the auto- 
matic equipment. This card function is frequently of double 
convenience, for the stencils are compact, stiff, and easily stored, 
and they can be automatically reproduced on standard machines. 
One of the early uses of punched cards as stencils was that by 
J. D. Elder; he revised and extended the factor stencils originally 
prepared by D. N. Lehmer 3 and punched the results on cards. 
This publication is out of print, but copies of the stencils can be 
prepared quickly and easily by means of a reproducing punch. 

9-2-2. Remington Rand Card. The Remington Rand Card 
has 90 vertical columns each with six punching positions. The 
card is divided into an upper and lower section, each section 
extending across the length of the card. There are 45 columns 
and six rows in each section; this is an adaptation of the Powers 
card which originally contained a total of 45 columns. The 
Remington Rand equipment will operate on cards with a capacity 
of 45 columns or 90 columns. Each column of the card is headed 
by a position, the five positions vertically beneath this repre- 
senting, respectively, (1,2), (3,4), (5,6), (7,8), and 9. The odd 
numbers and are represented by a single punch; the even 
numbers are represented by a two-hole punch in each column, 
one punch in the hole for the even number, the other in the nine 
position. For representing the alphabet, a two- or three-hole 
punch is required. 

This card is the basic unit for the operation of Remington Rand 
equipment. 

9-3. Input 

Data are introduced to punched cards through a manually 
operated punch with a keyboard similar to that of a typewriter. 



PUNCHED-CARD COMPUTING SYSTEMS 151 

This keyboard actuates a mechanism which translates a number, 
letter, or other code into the proper punched designation on 
the card. For the alphabet, a single key may operate a triple 
punch simultaneously. 

Data may also be introduced to cards by automatic transcrip- 
tion equipment or by punches operating from control panel or 
plugboard instructions. 

9-3-1. IBM Input. Data may be introduced to IBM cards 
by means of a manually operated punch, a tape-reading punch, 
or special equipment for automatically transcribing data on 
cards. Many models of manual punches are available. These 
punches may produce double punches for alphabetical informa- 
tion from a single stroke on a keyboard; they may print punched 
information at the top of the column; they may provide for 
automatic reproduction of punched cards. The larger punches 
have many other attachments to provide automatic aids to 
accuracy and convenience, such as tabulating stops, skip bars, 
automatic ejection and feed, etc. 

The tape reader automatically reads punched tape (like tele- 
type tape) and transcribes the data represented by the holes in 
the tape to a deck of cards. Other equipment can perform the 
reverse operation, reading the holes punched in the cards and 
producing a tape on which the same data are represented. 

With regard to input equipments, it may be of interest to 
note that the International Business Machines Corporation has 
designed and installed special equipment to read wind-tunnel 
data directly from various gauges and to transcribe these data on 
card punches. No standard equipment to accomplish this, 
however, has been advertised as yet. 

The existence of the Electric Punched Hole Verifier should be 
noted in connection with punch equipment. This device provides 
automatic detection of variation of entering information by sepa- 
rate operators. The original punched cards are set on the 
machine, and the verifier operator, reading data from the punched 
source, introduces these data to a keyboard similar to that on the 
punch. If the punching in a card column does not agree with the 
key struck on the verifier, an error light is illuminated, and oper- 
ation of the verifier is automatically stopped. If there are no 
errors, the verifier card is ejected and stacked. The verifier may 
be used to verify either alphabetical or numerical punches. 



152 COMPUTING SYSTEMS 

The manually operated punches and the verifier operate at 
typewriter speeds; manual punches with automatic reproducing 
attachments operate at about 10 columns per second. The tape 
reader operates at about 8 columns per second. Normally, 
manual punching of numerical data from clear manuscript may 
be carried out by an experienced operator at the rate of 125 cards 
per hour with 80 punches per card. 

9-3-2. Remington Rand Input. The Remington Rand equip- 
ment for the input of data is similar in character to the IBM 
input equipment. A typical input mechanism manufactured by 
Remington Rand is the 90-column Alphabetical Punch, which 
introduces both alphabetical and numerical information on a 
card. With this punch, a verifying attachment is available for 
preparing the punched cards for mechanical verification. The 
90-column Alphabetical Punch may include an attachment for 
serial numbering of each card which has passed through the 
machine. The serial number is printed along the left edge of 
the card; a counter on the front of the machine indicates the 
number of cards punched or stamped. 

The punch equipment for data input is manually operated and 
operates at typewriter speeds. Other equipments, such as the 
Summary Card Punch, the Reproducing Punch, and the Syn- 
chro-Matic Punch, operate at speeds above that of a typewriter. 
Their function is to read automatically data punched on cards 
in order to punch these data on other cards or to print them on 
paper tape. 

It is interesting to note that the Remington Rand punches 
contain an electrical link from the keyboard to the characters on 
the carriage. This is the only such linkage in the Remington 
Rand equipment. All verifying, reading, arithmetic operations, 
and printing in the Remington Rand equipment are actuated by 
mechanically sensing the card coding. This is claimed to be an 
advantage in favor of accuracy. The mechanical nature of these 
operations, however, prevents the use of orders presented to the 
machine by electrical plugboard connections, a function which 
gives the IBM equipment its wide application and flexibility. 
For introducing numerical data and various special orders, the 
use of punched tape, employed by IBM, has made that com- 
pany's equipment more readily adaptable to certain types of 
scientific calculation. 



PUNCHED-CARD COMPUTING SYSTEMS 



153 



9-4. Machines for Punched -card Computations 

Machines for performing punched-card computations are 
capable of carrying out the following operations: counting, sort- 
ing, consulting tables, addition, subtraction, multiplication, 
division, and the printing of results. A large number of com- 
putational techniques have been developed to utilize these opera- 
tions. The development of these techniques has, for example, 
enabled such operations as multiplication to be carried out on 
machines that are not advertised as multipliers. All the opera- 
tions enumerated above can be performed on IBM and on 
Remington Rand equipments. IBM equipments have been, 
in general, much more widely used on scientific calculation. 

9-4-1. IBM Machines. The principal IBM equipments 
useful in performing scientific calculations are the Sorter, the 
Collator, the Tabulator, the Electric Multiplier, and the Cal- 
culating Punch. These machines usually form a calculating 
unit when they are used together for scientific work. Table 9-2 
is presented as an outline of the operations which can be per- 
formed on each of these machines. Detailed information 

Table 9-2 



IBM machine 


Count 


Sort 


Consult 
tables 


Add 


Subtract 


Multiply 


Divide 


Sorter 


X 


X 


X 










Collator 




X 


X 










Tabulator 


X 






X 


X 






Electric Mul- 
tiplier 








X 


X 


X 




Calculating 
Punch 








X 


X 


X 


X 


Electronic 
Calculating 
Punch 


X 






X 


X 


X 


X 



concerning these operations is available in published form. 1 
In addition, the utilization of this equipment to carry out prob- 



154 COMPUTING SYSTEMS 

lems in scientific computing is described in many papers; a 
bibliography of some principal papers in English concerning 
the use of IBM machines in scientific research, statistics, and 
education has been published by the International Business 
Machines Corporation; it is available from them. With their 
permission, the sections entitled Scientific Research and Statis- 
tics are included in the Bibliography at the end of this chapter. 
For convenience, a brief description of the machines and of the 
operations will be given here in condensed form. 

The Sorter. The electric punched-card sorting machine is a 
machine which separates cards into 13 pockets depending upon 
the character punched in a chosen column. For normal numeri- 
cal sorts, one of the 12 pockets receives cards appropriate to 
one of 12 different punches, and the 13th receives cards with no 
punch at all in the chosen column. Cards already sorted on 
one column can be stacked together and resorted on another 
column so that, finally, a deck of cards can be ordered in accord- 
ance with numbers punched across several columns. If a column 
contains two or more punches, the Sorter normally selects in 
accordance with the punch lowest on the card in this column. 
However, it is possible to disable readers at any chosen level; 
by this means alphabetical sorting is easily brought about on two 
runs per column through the sorter. The sorter can be procured 
with 15 counters attached, so that the number of punches 
detected in each of the 12 column positions, the number of cards 
unpunched in the chosen column, a total, and a subtotal can be 
automatically counted. Counting may be done without sorting, 
if desired. Sorting can be performed at the rate of 450 cards per 
column per minute; the newest machines operate at 650. 

The Electric Punched-card Collator. This machine is designed 
to carry out somewhat more complicated rearrangements of 
cards than the sorter. It is particularly useful in connection 
with consulting tables of functions. The machine has two feeds 
and four output pockets. It achieves a considerable versatility 
in its disposition of the cards from the two input decks to the 
output pockets by virtue of a removable plugboard into which 
plug wires are manually inserted to provide the desired opera- 
tions. The operation of the machine depends upon comparison 
of the information read from three cards. The two leading cards 
from one feed and the one leading card from the other are read 



PUNCHED-CARD COMPUTING SYSTEMS 155 

simultaneously. Comparison circuits permit identification of 
several cases of relative order of the cards, and the various cases 
set off driving mechanisms to advance the cards from one of the 
two feeds, channeling the most advanced card from this feed into 
one of the four output pockets. As an example, a sorted set of 
arguments could be introduced to one feed and a function table 
(containing cards for arguments including those in the sorted 
set, this function table sorted according to argument) could be 
introduced in the other. The four pockets could be assigned to: 

1. Functional-value cards corresponding to arguments appear- 
ing in the first deck. 

2. Functional-value cards corresponding to arguments not 
appearing in the first deck. 

3. Cards from the first deck for which functional values were 
found. 

4. Cards in the first deck for which no functional values were 
found. 

(The fourth output pocket would contain cards which might have 
been misplaced in the original sorting process, for which the cards 
from the functional deck have been lost or otherwise removed, 
etc.) In a later operation the Collator can be used to merge the 
separated decks into their original orders. 

The Collator works at a speed of about 240 operations per 
minute. Each operation may involve the transfer of one or two 
cards, or none, to an output pocket, depending upon its nature. 
Almost all Collator operations are on decks which have already 
been sorted and which are to be dealt into the four pockets with 
some merging or separation but with no further reordering. 

The Tabulator. There are several types of tabulators, of which 
a standard useful model is the Type 416 Electric Punched Card 
Accounting machine. The company describes this machine 
as a combination adding, subtracting, and printing machine. 
Control of the Tabulator is flexible. It is brought about through 
wiring set on a removable plugboard. 

The input to the Tabulator is from a single feed with reading 
stations examining the two most advanced cards simultaneously. 
There are two reading stations: the control station and the arith- 
metic reading station. The control station causes the device 
to carry out automatically various arithmetical or printing opera- 



156 COMPUTING SYSTEMS 

tions, such as addition, subtraction, clearing, and printing. The 
arithmetical operations are all based on addition; access to the 
adding counters is only through the more advanced reading 
brush. 

The Tabulator, through its provision for automatic printing, 
is one of the most valuable of all the IBM devices for automatic 
computation. As pointed out earlier, the possibility of printing 
the output of a computation without any manual intervention is 
an important step in increasing the reliability of the output. It 
has become standard practice at many places to photograph auto- 
matically printed numbers for photo-offset printing to avoid 
typesetting errors. 

The output of the Tabulator can be punched on another deck 
of cards through use of the summary punch. Rearrangement of 
data on summary cards is possible by virtue of a plugboard on 
the punch. 

The Tabulator operates at a speed of 9,000 cards per hour 
while tabulating or listing. If its output is to be punched on 
another deck of cards, the operation is delayed by about 1.2 
seconds per card punched. 

The Electric Multiplier. This machine multiplies numbers 
punched on cards and punches the products. One factor is 
always derived from an individual card, and the other may be 
obtained either from a different field on the same card or from 
master cards containing group multipliers. The machine will 
accept eight-digit multiplicands and multipliers and produce 
products up to and including 16 digits. The Electric Multiplier 
contains a summary counter which will accumulate a 16-digit 
summary of products. 

The machine is available in a model which will add and sub- 
tract as many as three numbers punched on the same card. Two 
of these numbers may each contain 12 digits; the third may 
contain as many as eight digits. Some slightly more compli- 
cated operations can be performed combining addition with 
multiplication. 

The speed of the machine depends on the number of digits in 
the multiplier. Ten-digit products of two eight-digit factors 
can be accumulated at the rate of 870 per hour; if they are 
punched, however, the rate is slowed to 730 per hour. For the 
subtracting function, the speed is about 1,000 cards per hour. 



PUNCHED-CARD COMPUTING SYSTEMS 157 

The Calculating Punch. The calculating punch will multiply, 
divide, add, and subtract. It is capable also of carrying out a 
short sequence of operations automatically. The machine will 
multiply two 10-digit factors to yield a 19-digit product. It will 
accept a 20-digit dividend and a 10-digit divisor to yield a 10-digit 
quotient. The possibilities of combining operations in this 
machine are fairly great. The speed is controlled mainly by the 
time of punching; it also depends upon the size of the numbers 
involved. Twenty-digit products may be computed at the 
rate of 1,000 per hour. Division speed is considerably less than 
this; for a five-digit quotient, the speed is approximately 540 
cards per hour. There are two calculating punches, the Type 
602 and the Type 602A. The statements made above apply to 
both machines, but the 602A is a little more versatile. 

The Electronic Calculating Punch. During the early part of 
1948 the availability of the IBM Electronic Calculating Punch, 
Type 604, was announced. This machine performs all the 
basic arithmetic operations by electronic methods. The numer- 
ical instructions to the machine are read from IBM cards; 
the results of the calculations are punched automatically on 
cards. The Type 604 is much faster and is a more versatile 
machine than the Calculating Punch, Type 602. It is faster than 
the 602 A but not more versatile. 

The Electronic Calculating Punch is a combination of two 
separate units, a punching unit similar to an IBM Gang Punch 
and an electronic unit for performing the arithmetic operations. 
The punch receives operating instructions from two control 
panels; these panels control factor reading, result punching, 
checking, and gang punching. The electronic calculating unit 
has one control panel for instructions relating to the type of 
arithmetic operation to be performed and to the transfer of 
numbers from one storage unit to another. 

The Type 604 can perform 60 program steps, or operations, 
per card; a program step includes any one of the four arithmetic 
operations, or a number transfer. The time usually allotted 
to a group of program steps is about 80 milliseconds; this is the 
time required for a card to pass from the first reading station to 
the punching station. 

Numerical information read from cards may be held in the 
machine for as many arithmetic operations as may be required. 



158 



COMPUTING SYSTEMS 



For example, a constant factor may be stored in either the factor 
or the general storage for successive use. For calculating, the 
machine has a reading capacity of 21 digits per card. Since the 
capacity of the electronic part of Type 604 is 50 digits, the total 
number of digits available for punching is restricted to 29. 
It is possible to increase the reading capacity of the machine to 
37 digits if the number of digits to be punched is reduced to 13. 

Group operations of arithmetic processes may be performed 
with the Type 604; numerical factors may be accumulated from a 
group of cards and the final results punched on the last card of a 
group. Data may also be gang-punched on cards at the same 
time as results of arithmetic calculations are punched on the card. 
Numerical calculations and the punching of data on cards are 
automatically checked by the machine. 

9-4-2. Remington Rand Machines. The principal Remington 
Rand machines useful in performing scientific calculations are 
the Sorter, the Interfiling Reproducing Punch, the Tabulator, 
and the Printing Multiplier Punch. These machines all operate 
on a mechanical basis; the sensing mechanism is not a brush 
contact but is an actual mechanical contact. Table 9-3 repre- 
sents the distribution of functions among the several Remington 
Rand machines under consideration. The following is a brief 

Table 9-3 



Remington 
Rand Machine 


Count 


Sort 


Consult 
tables 


Add 


Subtract 


Multiply 


Divide 


Sorter 


X 


X 












Interfiling Re- 
producing 
Punch 


X 




X 










Tabulator 


X 






X 


X 






Printing Mul- 
tiplier Punch 








X 


X 


X 


X 



description of each of the machines listed in the table above: 

The Sorter. The Sorter mechanically selects and arranges 
cards in any required order. The cards may be selected and 
sorted into 12 receiving magazines. This selection is made by 



PUNCHED-CARD COMPUTING SYSTEMS 159 

using 12 selecting pins, one each for the 12 positions of the 
column. Each of the 12 receiving magazines has an individual 
counter with a capacity of 9,999. In addition, the sorter has a 
total counting register which counts the number of cards going 
into all the receiving magazines. The speed of the counting 
sorter is 420 cards per minute for 45-column cards, or 250 cards 
per minute for 90-column cards. Each receiving magazine has 
a capacity of 350 to 375 cards and the total feeding-magazine 
capacity is 600 cards. 

The Interfiling Reproducing Punch. The function of this 
machine is to feed two files of cards from two feeding magazines 
so as to (1) compare the two files, (2) punch the cards fed by the 
lower magazine subject to an established comparison with the 
cards in the upper feeding magazine, and (3) interfile selected 
cards from both files and segregate certain of the cards from both 
files according to an established comparison. There are four 
card-feeding magazines, two upper feeding magazines for pattern 
cards and two lower feeding magazines for cards to be punched. 
There are five card-receiving magazines, two upper magazines 
for the segregation of cards from the upper feeding magazine 
and two lower receiving magazines for segregating the cards 
from the lower feeding magazine. A third lower magazine is 
used for the interfiling of cards fed from both the upper and lower 
feeding magazines. 

The speed of operation of the Interfiling Reproducing Punch 
is 100 cards per minute. Each feeding magazine has a capacity 
of 600 cards; each receiving magazine has a capacity of 750 
cards. 

The Tabulator. The principal function of the Tabulator is to 
translate and print numerical and alphabetical information 
punched in tabulating cards. The Model 3 90-column Tabulator 
will perform automatically the operations of addition, subtrac- 
tion, and printing; it will also produce totals and/or grand totals. 
Its operation is fully automatic once the cards are inserted into 
the feeding mechanism. Although there are a number of models 
in manufacture, probably the one of interest scientifically is the 
Type 3100 Alphabetic Tabulator which has 100 printing sectors: 
50 sectors each equipped with 10 numerical characters and 26 
alphabetical characters, and 50 sectors each equipped to print 10 
numerical characters only. Such tabulators will operate at 



160 COMPUTING SYSTEMS 

100 cards per minute and have a feeding-magazine capacity of 
625 cards and a receiving-magazine capacity of 850 cards. 
Data are printed on a standard 20-inch carriage, the 100 sectors 
occupying a length of about 15% inches across the sheet. 

The Printing Multiplier Punch. This machine is capable of 
performing addition, subtraction, multiplication, and division 
with a capacity for printing totals of all products up to 14 digits. 
This machine has a capacity for either 45 or 90 columns inter- 
changeably; results are printed on paper tape or may be punched 
on cards. The punch is also equipped with both automatic and 
manual control. A keyboard allows an operator to set up 
manually one or both factors which are to be handled arith- 
metically or whose results are to be punched on cards or printed 
on the recording tape. The use of wiring units provides the 
machine with flexibility in setting up various types of problems. 
With this punch, it is, possible to print and punch information 
on a summary card, to print progressive and final totals, and to 
interject amounts to be subtracted from these totals or added to 
them with a printed registration of these operations being made. 
The speed of the machine for six-by-six-digit multiplication is 
1,200 calculations per hour. The product capacity is 12 digits, 
and the total capacity for printing is 14 digits per row. 

9-5. Computational Operations on IBM Machines 

On page 153, a table showing which IBM machines could be 
used to carry out each of seven basic operations was presented, 
and a brief description of these machines followed. Thus, Sec. 
9-4-1 analyzed Table 9-2 more or less horizontally. This 
section will give an equally brief outline of the table vertically, 
according to the operations. IBM equipments will be discussed 
here to illustrate these operations. 

9-5-1. Counting. The number of cards in which chosen 
punches appear in a particular column can most rapidly be deter- 
mined by running the cards through an ordinary counting sorter. 
For this the sorter would be set to retain the cards in their 
original order, and the 15 counter dials would give the number 
of cards for which each of 12 possible punches was detected, the 
number for which no punch was detected, the subtotal, and the 
total. The tabulator in the larger machines can be wired to 



PUNCHED-CARD COMPUTING SYSTEMS 161 

count, but this operation is uneconomical in time and in the use 
of expensive machinery. 

9-5-2. Sorting. The most rapid and convenient sorting of a 
badly disarranged deck is brought about through the use of the 
Sorter. In this device sorting is performed one column at a 
time at the rate of 450 cards per column per minute. A sort 
can be made on a number of digits by repeated use of this device. 
To do this, the initial sort is made on the lowest ranking digit. 
The cards are gathered from the pockets and stacked in order; 
they are then re-sorted on the next lowest ranking digit, etc. 
Thus, if 1,000 numbers from 000 through 999 were to be put in 
order with the lowest numbered card on top, the following 
sequence of operations would be carried out: 

1. Sort on the right-hand column and gather the cards with 
those in the zero pocket on top, those in the one pocket 
next, etc. 

2. Re-sort from the middle column and gather the cards with 
those in the zero pocket on top, the one pocket next, etc. 

3. Sort on the left column and gather the cards with those in 
the zero pocket on top, the one pocket next, etc. 

This completes the sort. The total sorting time will have been 
6% minutes for the three passages of 1,000 cards at the rate of 
450 cards per minute. For decks which are partially sorted, 
or for the merging of two sorted decks into a single deck in order, 
the Collator may be more efficient than the Sorter. The condi- 
tions under which efficient sorts can be made on the Collator are 
fairly apparent from the specifications on the machine, and no 
guidance to its use more than a familiarity with its operations 
seems necessary. 

Block sorting can be performed on the highest digit if sorting 
above a given digit is required. In this case, time is saved over 
the lowest digit sort, since there are fewer cards involved in the 
sorting operation. 

9-5-3. Table Consulting. To consult tables, the easiest 
procedure is to insert two sorted decks into the Collator. This 
procedure has already been outlined on pages 154-155 under the 
Electric Punched-card Collator. The Sorter, also, can be used 
for consulting tables. 



102 COMPUTING SYSTEMS 

9-5-4. Addition and Subtraction. Addition and subtraction 
are most economically carried out on the Tabulator, except 
where combined with multiplication or division or where results 
are recorded on the same card. In this case, the Calculating 
Punch may be profitably used. In this operation the Tabulator 
is a fairly versatile machine and is even capable of carrying out 
some multiplications efficiently. A simple automatic control 
program can be introduced through the control brushes of the 
tabulator, using either special control cards merged into the 
final deck through a Collator or the information on the original 
deck. 

9-5-5. Multiplication. Except for a few simple programs in 
multiplication which can be carried out effectively on the Tabu- 
lator, multiplication requires the use of either the Electric 
Multiplier or the Calculating Punch. In these devices the 
computation is carried out straightforwardly as described in 
connection with the descriptions of the machines. 

9-5-6. Division. Division must be carried out on the Cal- 
culating Punch unless a method of successive approximation is 
used. Various schemes of obtaining the reciprocal of a number 
without actually dividing may be applied to other machines. 
These schemes vary from table consulting, which may be carried 
out on the Collator, to use of a successive approximation by a 
method described in Chap. 7, which involves a somewhat similar 
table of reciprocals combined with a multiplication operation. 
Thus, it is true that well-equipped punched-card-machinery 
installations which do not contain a Calculating Punch may still 
carry out efficiently those computations involving division. 

9-6. Transcription Devices 

Transcription of results for reuse in other machines is brought 
about through reproducing or summary punches of various 
kinds. The use of the summary punch with the Tabulator has 
already been described above. Other punches permit rapid 
reproduction of a deck of cards with modification in the position 
of the punched information, with insertion of gang-punched 
information, representing several specific functions of these 
punches. One device, the Type 513 Electric Card Reproducing 
Punch, can be used as a gang punch, a reproducing punch, and 



PUNCHED-CARD COMPUTING SYSTEMS 163 

as a summary punch. Through temporary wiring on a 
standard plugboard, considerable rerouting of information is 
possible in this machine, together with enough rearrangement 
during the data transcription to permit efficient reentry into the 
other calculating equipments. 

This punch operates at a fixed speed of 100 cards per minute 
for reproducing and gang-punching operations. Summary 
punching from an accounting machine requires 1.2 seconds per 
card. 

This machine contains a comparing feature which provides 
assurance that the punched card agrees with the original from 
which it was punched. 

Other transcription machinery includes ordinary manual 
punches with duplicating attachments. These will reproduce a 
card already punched using the same equipment that is used in 
the manual punching. Some of these machines type the punched 
information across the top of the card for easy visual reading; 
this is a service which is not provided by the ordinary reproducing 
punch. On the other hand, their operation is much slower than 
the reproducing punch described earlier. 

The information punched on a card can be printed across 
the edge of the card by means of punched-card interpreters. 
Numeric machines will print the information, punched in any 
four to five columns, at the rate of 4,500 cards per hour. Another 
machine will print 60 characters, either numerical or alphabetical, 
at the rate of 3,600 cards per hour. 

Other transcription devices have already been mentioned. 
These include the printing tabulator and the machines for tran- 
scribing information on tape. 

9-7. Acquisition of Equipment 

Machines for punched-card computation are available on a 
lease or purchase basis from Remington Rand, Incorporated, 
and on lease from, but not generally for sale by, the International 
Business Machines Corporation. This corporation usually 
gives particular consideration to individual problems occurring 
at each proposed installation. The corporation maintains 
offices in most principal cities. 

For computation problems which are of too infrequent occur- 



164 COMPUTING SYSTEMS 

rence to justify the installation of an adequate punched-card 
facility, the International Business Machines Corporation main- 
tains complete machine services at several of its offices. Here 
machinery is available to carry out many computations, and the 
corporation accepts these on a job basis. Bids for such services 
will be made by the corporation on the basis of price per cards 
used. 

Tables and many mathematical functions are available in 
punched-card editions from several sources. Information con- 
cerning these tables may be obtained from the International 
Business Machines Corporation or the Watson Scientific Com- 
puting Laboratory. Information concerning other published 
tables of mathematical functions is available in periodical 
form. 3 

9-8. Mathematical Operations to Which IBM Equipment Is 
Applicable 

The International Business Machines Corporation has listed 
the following operations as typical of mathematical operations to 
which the equipment has been successfully applied: 

1. Construction of mathematical tables. 

2. Differencing, to any order desired. 

3. Interpolation, using first-, second-, and higher order 
differences. 

4. Numerical integration. 

5. Harmonic analysis. 

6. Solution of simultaneous equations. 

7. Summation of series. 

8. Multiplication of series. 

9. Matrix and determinant calculations. 

10. Method of least squares. 

11. Differential equations. 

12. Fourier analysis. 

13. Successive approximation methods. 

14. Curve fitting. 

More specific information concerning actual applications 
which have been made may be obtained by examining the 
accompanying bibliography 2 on the applications of punched-card 
machinery to scientific research and statistics. 



PUNCHED-CARD COMPUTING SYSTEMS 165 

9-9. IBM Card-programmed Electronic Calculator 

Since the foregoing sections of this chapter were written the 
International Business Machines Corporation has announced 
a new calculator which combines several IBM units in a sequence- 
controlled machine well adapted for scientific and engineering 
work. 

It uses the Type 402 or the Type 417 Accounting Machine as 
the master control unit and printer. The program for a cal- 
culation is contained on a deck of cards which are fed into this 
Accounting Machine. Each card carries an eight-digit instruc- 
tion which specifies the locations in storage of the factors to be 
operated upon, the operation to be performed, and the disposition 
to be made of the result. 

For internal storage, the calculator uses the 80 mechanical 
counters with which the accounting machine is normally equipped. 
A complete record of a calculation can be printed at the rate of 
150 lines per minute. 

Cards are used for external storage. A Calculator Punch, 
Type 521, is used to prepare these cards and also to record on 
cards the results of the calculations. A Supplemental Storage 
Unit, Type 941, provides storage for 16 10-digit signed numbers 
relayed to it from the Type 604 Electronic Calculator. 

The Type 604 Electronic Calculator is used to perform the 
various arithmetic operations. The manufacturer states that 
additions or subtractions are performed at a speed in excess of 
2,000 a second, multiplications or divisions at 86 a second, and 
that any of these arithmetic operations can be combined. 

The units of which this card-programmed computing system 
is composed may be disconnected from each other and used 
individually to perform their normal accounting operations. 

REFERENCES 

1. Eckert, W. J., Punched Card Methods in Scientific Computation (Thomas 
J. Watson Astronomical Computing Bureau, Columbia University, 1940). 

2. International Business Machines Corporation, Bibliography on the Use of 
IBM Machines in Science, Statistics and Education (International Busi- 
ness Machines Corporation, New York, 1950). This bibliography 
consists of technical articles that describe researches in the various fields 
of science, statistics, and education conducted with the aid of IBM 
machines, and machine methods useful for such research. It is a revision 



166 COMPUTING SYSTEMS 

of The Use of IBM Machines in Scientific Research, Statistics, and Educa- 
tion, which was published in 1947. Reproduction of Part A entitled 
Science and Part B entitled Statistics is made with the permission of the 
IBM Corporation. 

3. Lehmer, D. N., Factor Stencils (Carnegie Institution of Washington, 
Washington, D.C.; stencils revised and extended by J. D. Elder, 1939). 

4. National Academy of Sciences, Mathematical Tables and Other Aids to 
Computation. Quarterly publication of the National Research Council, 
Washington, D.C.; first publication, January, 1943. 

BIBLIOGRAPHY 
Science 

Aiken, Howard H., and Grace M. Hopper, The Automatic Sequence Con- 
trolled Calculator, Electrical Engineering, Vol. 65, No. 8-9, pp. 384-391 
(August-September, 1946); No. 10, pp. 449-454 (October, 1946); No. 11, 
pp. 522-528 (November, 1946). 

Akushsky, I. J., On Certain Schemes of the Numerical Harmonic Analysis, 
Comptes Rendus (Doklady) de I'Academie des Sciences de VURSS (N.S.), 
Vol. 52, pp. 471-474 (1946) (Russian). 

Akushsky, I. J., The Four-Counter Scheme of Solution of Dirichlet's Prob- 
lem by Means of Punched-Card Machines, Comptes Rendus (Doklady) de 
I'Academie des Sciences de I'URSS (N.S.), Vol. 54, No. 8, pp. 659-662 
(1946) (Russian). 

Akushsky, I. J., Mathematical Tables on Punched Cards. Functions of One 
Variable, Akademiia Nauk SSSR, Otdelenie Tekhnicheskikh Nauk, No. 11, 
pp. 1405-1454 (1947) (Russian). 

Akushsky, I. J., New Methods for the Calculation of a Sum of Products with 
a Tabulator, Doklady Akademiia Nauk SSSR (N.S.), Vol. 60, No. 1, pp. 
5-8 (1948) (Russian). 

Akushsky, I. J., Numerical Solution of the Dirichlet Equation with the Aid 
of Perforated Card Machines, Comptes Rendus (Doklady) de V Acade'mie des 
Sciences de I'URSS (N.S.), Vol. 52, No. 5, pp. 375-378 (1946) (Russian). 

Akushsky, I. J., On the Numerical Solution of Dirichlet Problem on Punched- 
Card Machines, Comptes Rendus (Doklady) de V Academie des Sciences de 
I'URSS (N.S.), Vol. 54, No. 9, pp. 755-758 (1946) (Russian). 

Akushsky, I. J., The Process of Diagonal Summation on a Tabulator and 
Some of Its Applications, Akademiia Nauk SSSR, Bulletin, Classe des 
Sciences Techniques, No. 5, pp. 475-495 (1947) (Russian). 

Akushsky, I. J., Some New Methods of Calculating Sums of Products on a 
Tabulator, Akademiia Nauk SSSR, Otdelenie Tekhnicheskikh Nauk, No. 8, 
pp. 1193-1228 (1948) (Russian). 

Alt, Franz L., Multiplication of Matrices, Mathematical Tables and Other 
Aids to Computation, Vol. 2, No. 13, pp. 12-13 (January, 1946). 

Atanasoff, J. V., and A. E. Brandt, Application of Punched Card Equipment 
to the Analysis of Complex Spectra, Journal of the Optical Society of 
America, Vol. 26, No. 2, pp. 83-88 (February, 1936). 



PUNCHED-CARD COMPUTING SYSTEMS 167 

Bailar, John C, Jr., Karl F. Heumann, and Edwin J. Seiferle, The Use of 
Punched Card Techniques in the Coding of Inorganic Compounds, Journal 
of Chemical Education, Vol. 25, No. 3, pp. 142-143, 176 (March, 1948). 

Beevers, C. A., A Machine for the Rapid Summation of Fourier Series, 
Proceedings of the Physical Society (London), Vol. 51, Part 4, pp. 660-663; 
discussion, pp. 664-667 (July 1, 1939). 

Bell, William D., Punched Card Techniques for the Solution of Simultaneous 
Equations and Other Matrix Operations, Proceedings, Scientific Computa- 
tion Forum, pp. 28-31 (International Business Machines Corporation, 
New York, 1948). 

Bell, William D., A Simplified Punch-Card Approach to the Solution of the 
Flutter Determinant, Journal of the Aeronautical Sciences, Vol. 15, No. 2, 
pp. 121-122 (February, 1948). 

Belzer, Jack, George Gamow, and Geoffrey Keller, Dynamics of Elliptical 
Galaxies, Proceedings, Scientific Computation Forum, pp. 67-69 (Interna- 
tional Business Machines Corporation, New York, 1948). 

Bergman, Stefan, Construction of a Complete Set of Solutions of a Linear 
Partial Differential Equation in Two Variables, by Use of Punch Card 
Machines, Quarterly of Applied Mathematics, Vol. 4, No. 3, pp. 233-245 
(October, 1946). 

Bergman, Stefan, Punch- Card Machine Methods Applied to the Solution of 
the Torsion Problem, Quarterly of Applied Mathematics, Vol. 5, No. 1, 
pp. 69-81 (April, 1947). 

Berkeley, Edmund C, Giant Brains, or Machines that Think (John Wiley 
and Sons, Inc., New York, 1949). 270 pp. 

Blanch, Gertrude, Differencing on the Type 405 Accounting Machine, 
Proceedings, Scientific Computation Forum, pp. 14-22 (International Busi- 
ness Machines Corporation, New York, 1948). 

Bower, Ernest Clare, On Subdividing Tables, Lick Observatory Bulletin, 
Vol. 16, No. 455, pp. 143-144, November, 1933. 

Bower, Ernest Clare, Systematic Subdivision of Tables, Lick Observatory 
Bulletin, Vol. 17, No. 467, pp. 65-74 (April, 1935). 

Bramble, Clinton C, Applications of Punched Card Equipment at the Naval 
Proving Ground, Proceedings, Scientific Computation Forum, pp. 99-100 
(International Business Machines Corporation, New York, 1948). 

Brinkley, Stuart, R., Jr., and Robert W. Smith, Jr., Calculation of the 
Equilibrium Composition of Systems of Many Constituents, Proceedings, 
Scientific Computation Forum, pp. 77-82 (International Business Machines 
Corporation, New York, 1948). 

Brouwer, Dirk, On the Accumulation of Errors in Numerical Integration, 
Astronomical Journal, Vol. 46, No. 16, pp. 149-153 (October, 1937). 

Brouwer, Dirk, The Reference System with a View to Planetary Dynamics, 
Annals of the New York Academy of Sciences, Vol. 42, Article 2, pp. 133-149 
(October 15, 1941). 

Comrie, L. J., The Application of Commercial Calculating Machines to 
Scientific Computing, Mathematical Tables and Other Aids to Computation, 
Vol. 2, No. 16, pp. 149-159 (October, 1946). 



168 COMPUTING SYSTEMS 

Comrie, L. J., The Application of the Hollerith Tabulating Machine to 
Brown's Tables of the Moon, Monthly Notices of the Royal Astronomical 
Society, Vol. 92, No. 7, pp. 694-707 (May, 1932). 

Comrie, L. J., On the Construction of Tables by Interpolation, Monthly 
Notices of the Royal Astronomical Society, Vol. 88, No. 6, pp. 506-523 
(April, 1928). 

Comrie, L. J., Interpolation and Allied Tables, Nautical Almanac for 1937 
(London). 

Comrie, L. J., Recent Progress in Scientific Computing, Journal of Scientific 
Instruments, Vol. 21, No. 8, pp. 129-135 (August, 1944). 

Comrie, L. J., Tables for Interpolation to Tenths and Fifths by the End 
Figure Process, Nautical Almanac for 1931, pp. 828-835 (London). 

Cox, E. G., R. J. J. H. Gillot, and G. A. Jeffrey, The Molecular Structure 
of Thiophthen from X-Ray Crystal Analysis, Acta Crystallographica, Vol. 
2, Part 6, pp. 356-363 (December, 1949). 

Cox, E. G., L. Gross, and G. A. Jeffrey, A Hollerith Punched-Card Method 
for the Evaluation of Electron Density in Crystal Structure Analysis, 
Proceedings of the Leeds Philosophical and Literary Society, Vol. 5, Part 1 , 
pp. 1-13 (December, 1947). 

Cox, E. G., L. Gross, and G. A. Jeffrey, A Hollerith Technique for Comput- 
ing Three- Dimensional Differential Fourier Syntheses in X-Ray Crystal- 
Structure Analysis, Acta Crystallographica, Vol. 2, Part 6, pp. 351-355 
(December, 1949). 

Cox, E. G., L. Gross, and G. A. Jeffrey, Use of Punched Card Tabulating 
Machines for Crystallographic Fourier Synthesis, Nature, Vol. 159, No. 
4039, pp. 433-434 (March 29, 1947). 

Culley, Frank L., Use of Accounting Machines for Mass Transformation 
from Geographic to Military-Grid Coordinates, Transactions of the 
American Geophysical Union (National Research Council, Washington, 
D.C., November, 1942), Part 2, pp. 190-197. 

Curtiss- Wright Corporation, Punched Card Method in Structural Calculations, 
Report No. V-241-S-4 (Curtiss- Wright Corporation, Airplane Division, 
Buffalo, N.Y., March 23, 1945). 

Domb, C, On Iterative Solutions of Algebraic Equations, Proceedings of the 
Cambridge Philosophical Society, Vol. 45, No. 2, pp. 237-240 (April, 1949). 

Donohue, Jerry, and Verner Schomaker, The Use of Punched Cards in 
Molecular Structure Determinations. III. Structure-Factor Calculations 
of X-Ray Crystallography, Acta Crystallographica, Vol. 2, Part 6, pp. 
344-347 (December, 1949). 

Dreyer, H. J., Solution of Systems of Linear Equations by Means of Punched- 
Card Machines, No. F-TS-1046-RE (Headquarters Air Materiel Com- 
mand, Wright Field, Dayton, Ohio, December, 1946). 13 pp. 

Duncan, David B., and John F. Kenney, On the Solution of Normal Equations 
and Related Topics (Edwards Bros., Inc., Ann Arbor Mich., 1946). 35 pp. 

Dunstan, Lyle A., The General Solution Method of Power Network Analysis, 
American Institute of Electrical Engineers Technical Paper 48-103 (Febru- 
ary, 1948). 19 pp. 

Dunstan, Lyle A., Machine Computation of Power Network Performance, 



PUNCHED-CARD COMPUTING SYSTEMS 169 

American Institute of Electrical Engineers Technical Paper 47-87 (Decem- 
ber, 1946). 33 pp. 

Dwyer, Paul S., The Solution of Simultaneous Equations, Psychometrika, 
Vol. 6, No. 2, pp. 101-129 (April, 1941). 

Dwyer, Paul S., The Use of Tables in the Form of Prepunched Cards, 
Proceedings of the Educational Research Forum; pp. 125-127 (International 
Business Machines Corporation, New York, August, 1940). 

Earthquake Calculations Are Made by Punched Card Machines, Engineering 
News Record, Vol. 141, No. 12, p. 110 (September 16, 1948). 

Eckert, W. J., Calculating Machines, Encyclopaedia Britannica, 13th Edition 
(1947), Vol. 4, pp. 548-554. 

Eckert, W. J., The Computation of Special Perturbations by the Punched 
Card Method, Astronomical Journal, Vol. 44, No. 20, pp. 177-182 
(October, 1935). 

Eckert, W. J., Electronic and Electromagnetic Measuring, Computing and 
Recording Devices, Centennial Symposia, December, 1946, Harvard 
Observatory Monographs, No. 7, pp. 169-178 (1948). 

Eckert, W. J., Electrons and Computation, Scientific Monthly, Vol. 67, No. 5, 
pp. 315-323 (November, 1948). 

Eckert, W. J., The IBM Pluggable Sequence Relay Calculator, Mathe- 
matical Tables and Other Aids to Computation, Vol. 3, No. 23, pp. 149-161 
(July, 1948). 

Eckert, W. J., Punched Card Methods in Scientific Computation, (Thomas J. 
Watson Astronomical Computing Bureau, Columbia University, 1940). 
136 pp. (New edition in preparation.) 

Eckert, W. J., Punched-Card Techniques and Their Applications to Scientific 
Problems, Journal of Chemical Education, Vol. 24, No. 2, pp. 54-57, 74 
(February, 1947). 

Eckert, W. J., and Ralph F. Haupt, The Printing of Mathematical Tables, 
Mathematical Tables and Other Aids to Computation, Vol. 2, No. 17, pp. 
197-202 (January, 1947). 

Feinstein, Lillian, and Martin Schwarzschild, Automatic Integration of 
Linear Second-Order Differential Equations by Means of Punched Card 
Machines, Review of Scientific Instruments, Vol. 12, No. 8, pp. 405-408 
(August, 1941). 

Fenn, George S., Programming and Using the Type 603-405 Combination 
Machine in the Solution of Differential Equations, Proceedings, Scientific 
Computation Forum, pp. 95-98 (International Business Machines Corpora- 
tion, New York, 1948). 

Ferber, Ben, Planning Engineering Calculations for IBM Equipment, 
Proceedings, Scientific Computation Forum, pp. 52-53 (International Busi- 
ness Machines Corporation, New York, 1948). 

Ferris, Lorna, Kanardy Taylor, J. W. Perry, and Maria E. W. Torok, 
Bibliography on the Uses of Punched Cards, (American Chemical Society, 
Washington, D.C., 1949). 40 pp. 

Fletcher, A. J., C. P. Miller, and L. Rosenhead, An Index of Mathematical 
Tables (McGraw-Hill Book Company, Inc., New York, 1946). 450 pp. 

Frost, Arthur A., The Approximate Solution of Schrodinger Equations by a 



170 COMPUTING SYSTEMS 

Least Squares Method, Journal of Chemical Physics, Vol. 10, No. 4, pp. 
240-245 (April, 1942). 

George, M. C, An Annotated Bibliography of Some Early Uses of Punched 
Cards in Meteorology and Climatology, Bulletin of the American Meteoro- 
logical Society, Vol. 26, No. 3, pp. 76-85 (March, 1945). 

Gordon, Norman J., Extracting Land-Use Data from Punched Cards, 
American City, Vol. 63, p. 149 (June, 1948). 

Grems, M. D., and J. S. Kasper, An Improved Punched-Card Method for 
Crystal Structure-Factor Calculations, Acta Crystallographica, Vol. 2, 
Part 6, pp. 347-351 (December, 1949). 

Grosch, H. R. J., Harmonic Analysis by the Use of Progressive Digiting, 
Proceedings of the Research Forum, pp. 81-84 (International Business 
Machines Corporation, New York, 1946). 

Grosch, H. R. J., Multiplication of Small Matrices (International Business 
Machines Corporation, New York, June 1, 1949). 3 pp. 

Grosch, H. R. J., The Use of Optimum Internal Mathematical Tables, 
Proceedings, Scientific Computation Forum, pp. 23-27 (International 
Business Machines Corporation, New York, 1948). 

Grosch, H. R. J. (Editor), Proceedings, Scientific Computation Forum 1948. 

Grossman, D. P., The Application of Punched Card Machines to the Solution 
of a System of Linear Algebraic Equations by the Iteration Method, 
Akademiia Nauk SSSR, Otdelenie Tekhnicheskikh Nauk, No. 8, pp. 1229- 
1238 (August, 1948) (Russian). 

Gull, Cloyd Dake, A Punched Card Method for the Bibliography, Abstract- 
ing, and Indexing of Chemical Literature, Journal of Chemical Education, 
Vol. 23, No. 10, pp. 500-507 (October, 1946). 

Hainer, R. M., and Gilbert W. King, The Asymmetric Rotor. V. An Analy- 
sis of the 3.7-m Band of H 2 S by Punched-Card Techniques, Journal of 
Chemical Physics, Vol. 15, No. 2, pp. 89-91 (February, 1947). 

Harrison, Joseph O., Jr., Piecewise Polynomial Approximation for Large- 
Scale Digital Calculators, Mathematical Tables and Other Aids to Computa- 
tion, Vol. 3, No. 26, pp. 400-407 (April, 1949). 

Hartkemeier, Harry Pelle, and Herman E. Miller, Obtaining Differences 
from Punched Cards, Journal of the American Statistical Association, Vol. 
37, No. 218, pp. 285-287 (June, 1942). 

Hartree, Douglas R., Calculating Instruments and Machines (The University 
of Illinois Press, Urbana, 111., 1949). 138 pp. 

Hartree, Douglas R., Notes on Iterative Processes, Proceedings of the Cam- 
bridge Philosophical Society, Vol. 45, No. 2, pp. 230-236 (April, 1949). 

Harvard Computation Laboratory, A Manual of Operation for the Automatic 
Sequence Controlled Calculator, The Annals of the Computation Laboratory 
of Harvard University, Vol. 1 (Harvard University Press, Cambridge, 
Mass., 1946). 561 pp. 

Harvard Computation Laboratory, Proceedings of a Symposium on Large- 
Scale Digital Calculating Machinery, The Annals of the Computation 
Laboratory of Harvard University, Vol. 16 (Harvard University Press, 
Cambridge, Mass., 1948). 302 pp. 

Harvard Computation Laboratory, Tables of the Bessel Functions of the 



PUNCHED-CARD COMPUTING SYSTEMS 171 

First Kind, The Annals of the Computation Laboratory of Harvard Univer- 
sity, Vols. 3-12 (Harvard University Press, Cambridge, Mass., 1947-1949). 

Harvard Computation Laboratory, Tables for the Design of Missiles, The 
Annals of the Computation Laboratory of Harvard University, Vol. 17 (Har- 
vard University Press, Cambridge, Mass., 1948). 226 pp. 

Harvard Computation Laboratory, Tables of Generalized Sine- and Cosine- 
Integral Functions, The Annals of the Computation Laboratory of Harvard 
University, Vols. 18-19 (Harvard University Press, Cambridge, Mass., 
1949). 

Harvard Computation Laboratory, Tables of Inverse Hyperbolic Functions, 
The Annals of the Computation Laboratory of Harvard University, Vol. 20 
(Harvard University Press, Cambridge, Mass., 1949). 

Harvard Computation Laboratory, Tables of the Modified Hankel Functions 
of Order One-Third and of Their Derivatives, The Annals of the Computa- 
tion Laboratory of Harvard University, Vol. 2 (Harvard University Press, 
Cambridge, Mass., 1945). 235 pp. 

Hausman, L. F. and M. Schwarzschild, Automatic Integration of Linear 
Sixth-Order Differential Equations by Means of Punched-Card Machines, 
Review of Scientific Instruments, Vol. 18, No. 12, pp. 877-883 (December, 
1947). 

d 2 P 
Herget, Paul, Integration of the Differential Equation -r-^ = P • F(r) Using 

the Type 601 Multiplying Punch, Proceedings, Scientific Computation 
Forum, pp. 39-41 (International Business Machines Corporation, New- 
York, 1948). 

Herget, Paul, Numerical Integration with Punched Cards, Astronomical 
Journal, Vol. 52, No. 4, pp. 115-117 (November, 1946). 

Herget, Paul, and G. M. Clemence, Optimum-Interval Punched-Card Tables, 
Mathematical Tables and Other Aids to Computation, Vol. 1, No. 6, pp. 173- 
176 (April, 1944). 

Hodgson, M. L., C. J. B. Clews, and W. Cochran, A Punched-Card Modifica- 
tion of the Beevers-Lipson Method of Fourier Synthesis, Acta Crystallo- 
graphica, Vol. 2, Part 2, pp. 113-116 (April, 1949). 

Hollander, Frederick H., Punched Card Calculating and Printing Methods in 
the Nautical Almanac Office, Proceedings, Scientific Computation Forum, 
pp. 83-94 (International Business Machines Corporation, New York, 
1948). 

Hollerith, H., The Electrical Tabulating Machine, Journal of the Royal 
Statistical Society, Vol. 57, Part 4, pp. 678-682 (December, 1894). 

Hollerith, H., An Electric Tabulating System, School of Mines Quarterly 
(Columbia University), Vol. 10, No. 3, pp. 238-255 (April, 1889). 

Horton, H. Burke, and R. Tynes Smith, III, A Direct Method for Producing 
Random Digits in Any Number System, Annals of Mathematical Statistics, 
Vol. 20, No. 1, pp. 82-90 (March, 1949). 

Hotelling, Harold, Some New Methods in Matrix Calculation, Annals of 
Mathematical Statistics, Vol. 14, No. 1, pp. 1-34 (March, 1943). 

Hotelling, Harold, Further Points on Matrix Calculation and Simultaneous 
Equations, Annals of Mathematical Statistics, Vol. 14, No. 4, pp. 440-441 
(December, 1943). 



172 COMPUTING SYSTEMS 

Howe, H. Herbert, Computing Magnetic Observatory Results with Punched 
Cards, U.S. Coast and Geodetic Survey, Document No. T-113 (July 6, 
1948). 

Hurd, Cuthbert C, IBM Applications in Industrial Statistics, Proceedings, 
Scientific Computation Forum, pp. 45-48 (International Business Machines 
Corporation, New York, 1948). 

International Business Machines Corporation, IBM Automatic Sequence Con- 
trolled Calculator (International Business Machines Corporation, New 
York, 1945). 6 pp. 

International Business Machines Corporation, IBM Selective Sequence Elec- 
tronic Calculator, Form 52-3927-0 (International Business Machines 
Corporation, New York, 1948). 16 pp. 

International Business Machines Corporation, IBM Survey Computer, 
Form 50-3885-0 (International Business Machines Corporation, New York, 
October, 1947). 7 pp. 

Jennings, Philip D., and George E. Quinan, The Use of Business Machines 
in Determining the Distribution of Load and Reactive Components in 
Power Line Networks, Transactions of the American Institute of Electrical 
Engineers, Vol. 65, pp. 1045-1046 (July, 1946). 

Johnson, Curt I., A New Approach to the Design of Dynamically Loaded 
Extension and Compression Springs, American Society of Mechanical Engi- 
neers, Paper No. 48-SA-23 (1948). 18 pp. 

Johnston, Herrick L., Lydia Savedoff and Jack Belzer, Contributions to the 
Thermodynamic Functions by a Planck- Einstein Oscillator in One Degree of 
Freedom (Office of Naval Research, Department of the Navy, Washington, 
D.C., July, 1949). 159 pp. 

Kimball, Everett, Jr., A Fundamental Punched Card Method for Technical 
Computations (Bureau of the Census, Washington, D.C., undated). 12 pp. 

Kimball, Everett, Jr., A Method of Technical Computations by Punched Card 
Equipment (Bureau of the Census, Washington, D.C., undated). 8 pp. 

King, Gilbert W., Application of Punched Cards in Physical Chemistry, 
Proceedings, Scientific Computation Forum, pp. 70-73 (International Busi- 
ness Machines Corporation, New York, 1948). 

King, Gilbert W., Some Applications of Punched-Card Methods in Research 
Problems in Chemical Physics, Journal of Chemical Education, Vol. 24, 
No. 2, pp. 61-64 (February, 1947). 

King, Gilbert W., The Asymmetric Rotor. IV. An Analysis of the 8.5-/* 
Band of D 2 by Punched Card Techniques, Journal of Chemical Physics, 
Vol. 15, No. 2, pp. 85-88 (February, 1947). 

King, Gilbert W., A Method of Plotting on Standard IBM Equipment, 
Mathematical Tables and Other Aids to Computation, Vol. 3, No. 25, pp. 
352-355 (January, 1949). 

King, Gilbert W., Punched Card Tables of the Exponential Function, Review 
of Scientific Instruments, Vol. 15, No. 12, pp. 349-350 (December, 1944). 

King, Gilbert W., Paul C. Cross, and George B. Thomas, The Asymmetric 
Rotor. III. Punched-Card Methods of Constructing Band Spectra, 
Journal of Chemical Physics, Vol. 14, No. 1, pp. 35-42 (January, 1946). 

Kintas, John, A Survey of the IBM Project at Beech Aircraft Corporation, 



PUNCHED-CARD COMPUTING SYSTEMS 173 

Proceedings, Scientific Computation Forum, pp. 54-59 (International Busi- 
ness Machines Corporation, New York, 1948). 

Knudsen, Lila F., A Punched Card Technique to Obtain Coefficients of 
Orthogonal Polynomials, Journal of the American Statistical Association, 
Vol. 37, No. 220, pp. 496-506 (December, 1942). 

Kormes, Jennie P., and Mark Kormes, Numerical Solution of Initial Value 
Problems by Means of Punched-Card Machines, Review of Scientific Instru- 
ments, Vol. 16, No. 1, pp. 7-9 (January, 1945). 

Kormes, Mark, Numerical Solution of the Boundary Value Problem for the 
Potential Equation by Means of Punched Cards, Review of Scientific 
Instruments, Vol. 14, No. 8, pp. 248-250 (August, 1943). 

Kraft, Hans, Aerodynamic Lattice Calculations Using Punched Cards, 
Proceedings, Scientific Computation Forum, pp. 60-66 (International 
Business Machines Corporation, New York, 1948). 

Laderman, Jack, and Milton Abramowitz, Application of Machines to Dif- 
ferencing of Tables, Journal of the American Statistical Association, Vol. 41, 
pp. 233-237 (June, 1946). 

Leavens, Dickson H., The Digit or Group Extension Method of Multiplication 
(Harvard Business School, BS475, undated). 6 pp. 

Lehmer, D. N., Factor Stencils, revised and extended by J. D. Elder (Car- 
negie Institution, Washington, D.C., 1939). 27 pp. + 2135 stencils. 

Leppert, E. L., Jr., An Application of IBM Machines to the Solution of the 
Flutter Determinant, Journal of the Aeronautical Sciences, Vol. 14, No. 3, 
pp. 171-174 (March, 1947). 

Levin, Joseph H., Use of the IBM Relay Calculators for Technical Calcula- 
tions at Aberdeen Proving Ground, Proceedings, Scientific Computation 
Forum, pp. 101-104 (International Business Machines Corporation, New 
York, 1948). 

Lindberger, N. Arne, Integration of Second Order Linear Differential Equa- 
tions on the Type 602 Calculating Punch, Proceedings, Scientific Computa- 
tion Forum, pp. 34-38 (International Business Machines Corporation, New 
York, 1948). 

Lorant, Ralph R., Evaluation of Determinants with Complex Elements by 
Means of the IBM Punched Card Method, (International Business Machines 
Corporation, New York, June 1, 1946). 16 pp. 

Lynch, J., and C. E. Johnson, Programming Principles for the IBM Relay 
Calculators, Report No. 705 (Ballistic Research Laboratories, Aberdeen 
Proving Ground, Aberdeen, Md., October, 1949). 95 pp. 

McPherson, J. C, Mathematical Operations with Punched Cards, Journal 
of the American Statistical Association, Vol. 37, No. 218, pp. 275-281 
(June, 1942). 

McPherson, J. C, On Mechanical Tabulation of Polynomials, Annals of 
Mathematical Statistics, Vol. 12, No. 3, pp. 317-327 (September, 1941). 

McPherson, J. C, Solution of Simultaneous Equations by Punched Card 
Methods (International Business Machines Corporation, New York, 
undated). 5 pp. 

Maginniss, Frank J., Some Engineering Applications of IBM Equipment at 
the General Electric Company, Proceedings, Scientific Computation Forum, 



174 COMPUTING SYSTEMS 

pp. 49-51 (International Business Machines Corporation, New York, 
1948). 

Martin, Louis D., A Simplified Fine-Pitch Worm Gearing Standard, Paper 
No. 48-A-82 (American Society of Mechanical Engineers, New York, 
1948). 14 pp. 

Massachusetts Institute of Technology, Department of Electrical Engineer- 
ing, Center of Analysis, Tables of Supersonic Flow Around Cones, Technical 
Report No. 1 (Cambridge, Mass., 1947). 555 pp. 

Massachusetts Institute of Technology, Department of Electrical Engineer- 
ing, Center of Analysis, Tables of Supersonic Flow Around Yawing Cones, 
Technical Report No. 3 (Cambridge, Mass., 1947). 321 pp. 

Massachusetts Institute of Technology, Department of Electrical Engineer- 
ing, Center of Analysis, Tables of Supersonic Flow Around Cones of Large 
Yaw, Technical Report No. 5 (Cambridge, Mass., 1949). 125 pp. 

Mayor, Yoland, Applying Statistical Machines to Bibliographical Investi- 
gations, Chimie et Industrie, Vol. 35, No. 2, pp. 458-466 (February, 1936) 
(French). 

Milne, William Edmund, Numerical Calculus (Princeton University Press, 
Princeton, N.J., 1949). 393 pp. 

Morgan, John A., and D. E. H. Frear, Coding and Sorting Chemical Com- 
pounds by Means of Punched Cards, Journal of Chemical Education, 
Vol. 24, No. 2, pp. 58-61 (February, 1947). 

Mulligan, J. H., Jr., The Effect of Pole and Zero Locations on the Transient 
Response of Linear Dynamic Systems, Proceedings of the Institute of 
Radio Engineers, Vol. 37, No. 5, pp. 516-529 (May, 1949); No. 10, p. 1181 
(October, 1949). 

Munk, Max M., Use of Automatic Calculation Methods in Aeronautical 
Research, Aero Digest, Vol. 50, No. 5, pp. 68-69, 192 (September 1, 1945). 

Murray, Francis J., Simultaneous Linear Equations, Proceedings, Scientific 
Computation Forum, pp. 105-106 (International Business Machines 
•Corporation, New York, 1948). 

National Bureau of Standards, Bibliography on Automatic Digital Comput- 
ing Machinery (U. S. Department of Commerce, Washington, D.C., 
January, 1948). 10 pp. 

Nautical Almanac Office, The American Air Almanac, (U.S. Government 
Printing Office, Washington, D.C., January- April, 1942). 240 pp. 

Nautical Almanac Office, Tables of Sunrise, Sunset and Twilight, Supple- 
ment to the American Ephemeris, 1946 (U.S. Government Printing Office, 
Washington, D.C., 1945). 196 pp. 

Nowacki, W., Die Verwendung von Lochkartenmaschinen zur Fourier- und 
Patterson-Synthese von Kristallen, Chimia, Vol. 2, No. 12, pp. 3-7 (May 
21, 1948) (German). 

Oakley, Margaret B., and George E. Kimball, Punched Card Calculation of 
Resonance Energies, Journal of Chemical Physics, Vol. 17, No. 8, pp. 706- 
717 (August, 1949). 

Peakes, Gilbert L., Report Indexing by Punch Cards, Journal of Chemical 
Education, Vol. 26, No. 3, pp. 139-146 (March, 1949). 

Peebles, Glenn H., A Method for Calculating Airfoil Sections from Specifica- 



PUNCHED-CARD COMPUTING SYSTEMS 175 

tions on the Pressure Distributions, Journal of the Aeronautical Sciences, 
Vol. 14, No. 8, pp. 451-456 (August, 1947). 

Polachek, Harry, Computation of Shock Wave Refraction on the Selective 
Sequence Electronic Calculator, Proceedings, Scientific Computation Forum, 
pp. 107-122 (International Business Machines Corporation, New York, 
1948). 

Rankin, A. W., Calculation of the Multiple-Span Critical Speeds of Flexible 
Shafts by Means of Punched-Card Machines, Paper No. 45-A-46 (Turbine 
Generator Engineering Division, General Electric Corporation, Schenec- 
tady, N.Y., undated). 10 pp. 

Renner, H. W., Solving Simultaneous Equations through the Use of IBM Elec- 
tric Punched Card Accounting Machines (International Business Machines 
Corporation, Endicott, N.Y., 1946). 6 pp. 

Reynolds, William A., A Prepunched Master Deck for the Computation of 
Square Roots on IBM Electrical Accounting Equipment, Psychometrika, 
Vol. 11, No. 4, pp. 223-237 (1946). 

Robinson, Raphael M., Stencils for Solving x 2 = a (mod m) (University of 
California Press, Berkeley and Los Angeles, 1940). 14 pp. + 272 stencils. 

Sandomire, Marion M., Accumulating Cubes with Punch Cards, Journal of 
the American Statistical Association, Vol. 36, No. 216, pp. 507-514 (Decem- 
ber, 1941). 

Savedoff, Lydia G., Jack Belzer, and Herrick L. Johnston, Application of 
Punched Card Methods to the Computation of Thermodynamic Properties 
of Gases from Spectra, Proceedings, Scientific Computation Forum, pp. 
74-76 (International Business Machines Corporation, New York, 1948). 

Savedoff, Lydia G., Jack Belzer, and Herrick L. Johnston, Computation of 
Thermodynamic Properties of Gases from Spectra with the Aid of Punched 
Card Equipment, Volume of the Third Symposium on Combustion, Flame 
and Explosion Phenomena, pp. 595-599 (University of Wisconsin, Mad- 
ison, Wis., September, 1948). 

Scarborough, James Blaine, Numerical Mathematical Analysis (Johns 
Hopkins Press, Baltimore, Md., 1930). 416 pp. 

Schwarzschild, M., On Stellar Rotation. II, Astrophysical Journal, Vol. 106, 
No. 3, pp. 427-456 (November, 1947). 

Shaffer, P. A., Jr., and Linus Pauling, The Use of Punched Cards in Molecu- 
lar Structure Determinations. I. Crystal Structure Calculations, Journal 
of Chemical Physics, Vol. 14, No. 11, pp. 648-658 (November, 1946). 

Shaffer, P. A., Jr., Verner Schomaker, and Linus Pauling, The Use of Punched 
Cards in Molecular Structure Determinations. II. Electron Diffraction 
Calculations, Journal of Chemical Physics, Vol. 14, No. 11, pp. 659-664 
(November, 1946). 

Shapley, A. H., The Recurrence Tendency and Forecasts of Magnetic 
Activity, Transactions of the American Geophysical Union, Vol. 28, No. 5, 
pp. 715-721 (October, 1947). 

Shimizu, Tatsujiro, and Yoichi Katayama, Solutions of Non-Linear Equa- 
tions by Punched-Card Methods, Mathematical Japonicae, Vol. 1, No. 2, 
92-97 (1948). 

Snedecor, George W., Uses of Punched Card Equipment in Mathematics, 



176 COMPUTING SYSTEMS 

American Mathematical Monthly, Vol. 35, No. 4, pp. 161-169 (April, 1928). 
Sprinkle, Leland W., Two Numerical Methods of Integration Using Pre- 
determined Factors, Proceedings, Scientific Computation Forum, pp. 32-33 

(International Business Machines Corporation, New York, 1948). 
Targoff, Walter P., The Associated Matrices of Bending and Coupled 

Bending-Torsion Vibrations, Journal of the Aeronautical Sciences, Vol. 14, 

No. 10, pp. 579-582 (October, 1947). 
Targoff, Walter P., A Tabulation Method for the Calculation of the Critical 

Speed of Wing Divergence, Journal of the Aeronautical Sciences, Vol. 14, 

No. 8, pp. 487-488 (August, 1947). 
Thomas, George B., and Gilbert W. King, Preparation of Punched-Card 

Tables of Logarithms, Review of Scientific Instruments, Vol. 15, No. 12, 

p. 350 (December, 1944). 
Thomas, L. H., Computation of Statistical Fields for* Atoms and Ions, 

Proceedings, Scientific Computation Forum, pp. 123-128 (International 

Business Machines Corporation, New York, 1948). 
Tucker, Ledyard R., Factorial Rotation of Axes with Graphs Made by Machine, 

Bibliofilm Manuscript (University of Chicago Library, 1941). 
Tucker, Ledyard R., A Matrix Multiplier, Psychometrika, Vol. 5, No. 4, 

pp. 289-294 (December, 1940). 
Verzuh, Frank M., Evaluation of Higher Order Differences on the Type 602 

Calculating Punch, Proceedings, Scientific Computation Forum, pp. 9-13 

(International Business Machines Corporation, New York, 1948). 
Verzuh, Frank M., The Solution of Simultaneous Linear Equations with the 

Aid of the 602 Calculating Punch, Mathematical Tables and Other Aids to 

Computation, Vol. 3, No. 27, pp. 453-462 (July, 1949). 
Von Neumann, John, and H. H. Goldstine, Numerical Inverting of Matrices 

of High Order, Bulletin of the American Mathematical Society, Vol. 53, No. 

11, pp. 1021-1099 (November, 1947). 
Walker, Robert M., An Analogoue Computer for the Solution of Linear 

Simultaneous Equations, Proceedings of the Institute of Radio Engineers, 

Vol. 37, No. 12, pp. 1467-1473 (December, 1949). 
Whittaker, Edmund, and G. Robinson, The Calculus of Observations (Blackie 

& Son, Ltd., London, 1946). 397 pp. 
Whitten, C. A., Air-Line Distances Between Cities in the United States, 

Special Publication No. 238 (U.S. Government Printing Office, Washing- 
ton, D.C., 1947). 246 pp. 
Whitten, C. A., Triangulation- Adjustment by International Business 

Machines, American Geophysical Union Transactions of 1943, Part 1, 

pp. 31-32 (National Research Council, Washington, D.C., October, 1943). 
Womersley, J. R., Scientific Computing in Great Britain, Mathematical 

Tables and Other Aids to Computation, Vol. 2, No. 15, pp. 110-117 (July, 

1946). 

Statistics 

Bartlett, Neil R., Punched Card Technique for Computing Means, Standard 
Deviations, and the Product Moment Correlation Coefficient, and for List- 
ing Scattergrams, Science, Vol. 104, No. 2703, pp. 374-375 (October 18, 
1946). 



PUNCHED-CARD COMPUTING SYSTEMS 177 

Benjamin, Kurt, An IBM Technique for the Computation of Src 2 and Sxy, 
Psychometrika, Vol. 10, No. 1, pp. 61-67 (March, 1945). 

Benjamin, Kurt, Problems of Multiple-Punching with Hollerith Machines, 
Journal of the American Statistical Association, Vol. 42, No. 237, pp. 46-71 
(March, 1947). 

Benjamin, Kurt, Statistical Procedures in the Hospital, Proceedings, Hospital 
Financial Control Institute, pp. 58-63 (International Business Machines 
Corporation, New York, August, 1948). 

Berkson, Joseph, A Punch Card Designed to Contain Written Data and 
Coding, Journal of the American Statistical Association, Vol. 36, No. 216, 
pp. 535-538 (December, 1941). 

Berkson, Joseph, A System of Codification of Medical Diagnoses for Applica- 
tion to Punch Cards with Plan of Operation, American Journal of Public 
Health, Vol. 26, pp. 606-612 (June, 1936). 

Black, Bertram J., and Edward B. Olds, A Punched Card Method for 
Presenting, Analyzing and Comparing Many Series of Statistics for Areas, 
Journal of the American Statistical Association, Vol. 41, No. 235, pp. 347- 
355 (September, 1946). 

Black-Schaffer, Bernard, and Paul D. Rosahn, Methods of Analysis of Yale 
Autopsy Protocols, Including a Code for the Punched Card Study of 
Syphilis, Yale Journal of Biology and Medicine, Vol. 15, pp. 575-586 
(March, 1943). 

Brandt, A. E., Use of Machine Factoring in Multiple Correlation, Journal of 
the American Statistical Association, Vol. 23, No. 163, pp. 291-295 (Septem- 
ber, 1928). 

Caddy, E. Reid, Punch Cards Simplify Statistics, Modern Hospital, Vol. 52, 
No. 5, pp. 82-84 (May, 1939). 

Castore, George F., and William S. Dye, III, A Simplified Punch Card 
Method of Determining Sums of Squares and Sums of Products, Psycho- 
metrika, Vol. 14, No. 3, pp. 243-250 (September, 1949). 

Cochran, W. J., Use of IBM Equipment in an Investigation of the "Trun- 
cated Normal" Problem, Proceedings of the Research Forum, pp. 40-44 
(International Business Machines Corporation, New York, August, 1946). 

Comrie, L. J., G. B. Hey, and H. G. Hudson, Application of Hollerith Equip- 
ment to an Agricultural Investigation, Supplement to the Journal of the 
Royal Statistical Society, Vol. 4, No. 2, pp. 210-224 (1937). 

Deming, W. Edwards, and Leon Geoffrey, On Sample Inspection in the Proc- 
essing of Census Returns, Journal of the American Statistical Association, 
Vol. 36, No. 215, pp. 351-360 (September, 1941). 

Deming, W. Edwards, Benjamin J. Tepping, and Leon Geoffrey, Errors in 
Card Punching, Journal of the American Statistical Association, Vol. 37, 
No. 220, pp. 525-536, December, 1942. 

DuBois, Philip H., A Note on the Computation of Biserial r in Item Valida- 
tion, Psychometrika, Vol. 7, No. 2, pp. 143-146 (June, 1942). 

DuBois, Philip H., Some Statistical Operations on the Counting Sorter, 
Psychometrika, Vol. 6, No. 6, pp. 383-390 (December, 1941). 

Dunlap, Jack W., The Computation of Means, Standard Deviations and 
Correlations by the Tabulator when the Numbers are Both Positive and 



178 COMPUTING SYSTEMS 

Negative, Proceedings of the Educational Research Forum, pp. 16-19 
(International Business Machines Corporation, New York, August, 1940). 

Dunlap, Jack W., Note on the Computation of Tetrachoric Correlations, 
Psychometrika, Vol. 5, No. 2, pp. 137-140 (June, 1940). 

Dunn, Halbert L., Adaptation of New Geometric Code to Multiple Punch- 
ing in Mechanical Tabulation, Journal of the American Statistical Associa- 
tion, Vol. 27, No. 179, pp. 279-286 (September, 1932). 

Dunn, Halbert L, and Reed Rockwood, A Record System Suitable for Both 
Clinical and Statistical Medicine, Archives of Internal Medicine, Vol. 41, 
pp. 499-535 (April, 1928). 

Dunn, Halbert L., and Leonard Townsend, Application of Punched Card 
Methods to Hospital Statistics, Journal of the American Statistical Associa- 
tion, Vol. 30, No. 189A, pp. 244-248 (March, 1935). 

Dwyer, Paul S., The Computation of Moments with the Use of Cumulative 
Totals, Annals of Mathematical Statistics, Vol. 9, No. 4, pp. 288-304 
(December, 1938). 

Dwyer, Paul S., Recent Developments in Correlation Technique, Journal of 
the American Statistical Association, Vol. 37, No. 220, pp. 441-460 (Decem- 
ber, 1942). 

Dwyer, Paul S., Simultaneous Computation of Correlation Coefficients with 
Missing Variates, Proceedings of the Research Forum, pp. 20-27 (Interna- 
tional Business Machines Corporation, New York, August, 1946). 

Dwyer, Paul S., Summary of Problems in the Computation of Statistical 
Constants with Tabulating and Sorting Machines, Proceedings of the 
Educational Research Forum, pp. 20-28 (International Business Machines 
Corporation, New York, August, 1940). 

Dwyer, Paul S., and Alan D. Meacham, The Preparation of Correlation 
Tables on a Tabulator Equipped with Digit Selection, Journal of the 
American Statistical Association, Vol. 32, No. 200, pp. 654-662 (December, 
1937). 

Eckler, A. Ross, Some Elementary Machine Problems in the Sampling Work 
of the Census, Proceedings, Scientific Computation Forum, pp. 42-44 
(International Business Machines Corporation, New York, 1948). 

Edwards, Thomas I., The Coding and Tabulation of Medical and Research 
Data for Statistical Analysis, Public Health Reports, Vol. 57, No. 1, pp. 
7-20 (January, 1942). 

Ellis, Max E., and Arthur J. Riopelle, An Efficient Punched-Card Method of 
Computing 2x, Sx 2 , Sxy and Higher Moments, Psychometrika, Vol. 13, 
No. 2, pp. 79-85 (June, 1948). 

Gibbs, Frederic A., and A. M. Grass, Frequency Analysis of Electroenceph- 
alograms, Science, Vol. 105, No. 2718 pp. 132-134 (January 31, 1947). 

Hall, D. H., E. L. Welker, and Isabelle Crawford, Factor Analysis Calcula- 
tions by Tabulating Machines, Psychometrika, Vol. 10, No. 2, pp. 93-125 
(June, 1945). 

Harman, Harry H., Four Aspects of Factor Analysis, Proceedings of the 
Educational Research Forum, pp. 60-67 (International Business Machines 
Corporation, New York, August, 1940). 



PUNCHED-CARD COMPUTING SYSTEMS 179 

Hartkemeier, Harry Pelle, Principles of Punch-Card Machine Operation 
(The Thomas Y. Crowell Company, New York, 1942). 269 pp. 

Hartley, H. O., The Application of Some Commercial Calculating Machines 
to Certain Statistical Calculations, Supplement to the Journal of the Royal 
Statistical Society, Vol. 8, No. 2, pp. 154-173; discussion, pp. 173-183 
(1946). 

Hey, G. B., A New Method of Experimental Sampling Illustrated on Certain 
Non-Normal Populations, Biometrika, Vol. 30, Part 1, pp. 68-80 (June, 
1938). 

Hodges, Fred Jenner, Isadore Lampe, and Kurt Benjamin, Organized 
Clinical Investigation of Cancer, Supplement to University Hospital Bulletin 
(University of Michigan, University Hospital, February, 1948). 

Homeyer, Paul G., Mary A. Clem, and Walter T. Federer, Punched Card and 
Calculating Machine Methods for Analyzing Lattice Experiments Includ- 
ing Lattice Squares and the Cubic Lattice, Research Bulletin 347, (Iowa 
State College, Ames, Iowa, April, 1947). 

International Business Machines Corporation, Proceedings of the Educational 
Research Forum (International Business Machines Corporation, New York, 
August, 1940). 127 pp. 

International Business Machines Corporation, Proceedings of the Research 
Forum (International Business Machines Corporation, New York, August, 
1946). 94 pp. 

Kempthorne, O., The Analysis of a Series of Experiments by the Use of 
Punched Cards, Supplement to the Journal of the Royal Statistical Society, 
Vol. 8, No. 1, pp. 118-127 (1946). 

Kempthorne, O., The Use of a Punched-Card System for the Analysis of 
Survey Data, with Special Reference to the Analysis of the National Farm 
Survey, Journal of the Royal Statistical Society, Vol. 109, Part 3, pp. 284- 
295 (1946). 

Lorge, Irving, The Computation of the Hotelling Canonical Correlation, 
Proceedings of the Educational Research Forum, pp. 68-74 (International 
Business Machines Corporation, New York, August, 1940). 

Martin, Caroline R., Hospital Medical Statistics (J. B. Lippincott Company, 
Philadelphia, 1933). 89 pp. 

Martin, Caroline R., The Value of the Work of a Central Medical Statistical 
Bureau. Paper read before the American Hospital Association, Mil- 
waukee, Sept. 13, 1933. 7 pp. 

Meacham, Alan D., The Tabulating Machine Method as Used for Modern 
Mathematical Statistics (Edwards Bros., Inc., Ann Arbor, Mich., 1930). 
17 pp. 

Meacham, Alan D., The Value of the Collator in Using Prepunched Cards for 
Obtaining Moments and Product Moments, Proceedings of the Educational 
Research Forum, pp. 9-15 (International Business Machines Corporation, 
New York, August, 1940). 

Mendenhall, R. M., and Richard Warren, Computing Statistical Coefficients 
from Punched Cards, Journal of Educational Psychology, Vol. 21, No. 1, 
pp. 53-62 (January, 1930). 



180 COMPUTING SYSTEMS 

Millang, Amy, and G. F. Sprague, The Use of Punched Card Equipment in 

Predicting the Performance of Corn Double Crosses, American Society of 

Agronomy Journal, Vol. 32, No. 10, pp. 815-816 (October, 1940). 
Miller, Richard D., Punch Card Simplifies Records, Modern Hospital, Vol. 

55, No. 1, p. 45 (July, 1940). 
Milliman, Wendell A., Mechanical Multiplication by the Use of Tabulating 

Machines, Transactions of the Actuarial Society of America, Vol. 35, Part 2, 

pp. 253-264 (October, 1934); Vol. 36, Part 1, pp. 77-84 (May, 1935). 
Pearl, Raymond, Modern Methods in Handling Hospital Statistics, Johns 

Hopkins Hospital Bulletin, Vol. 32, No. 364, pp. 184-194 (June, 1921). 
Reagh, Russell R., A Simplified Code for Multiple Card Punching, Journal 

of the American Statistical Association, Vol. 29, No. 186, pp. 182-183 

(June, 1934). 
Reid, Mabel, Use of IBM Punch Cards, Public Health Nursing, Vol. 39, No. 

2, pp. 83-85 (February, 1947). 
Rosahn, Paul D., The Punched Card Technique as Applied to Autopsy 

Protocol Analysis, Journal of Technical Methods, Vol. 18, pp. 32-36 

(November, 1938). 
Royer, Elmer B., A Machine Method for Computing the Biserial Correlation 

Coefficient in Item Validation, Psychometrika, Vol. 6, No. 1, pp. 55-59 

(February, 1941). 
Royer, Elmer B., and Herbert A. Toops, The Statistics of Geometrically 

Coded Scores, Journal of the American Statistical Association, Vol. 28, No. 

182, pp. 192-198 (June, 1933). 
Schreiber, Frederic, and Aage Nielsen, Punch Card Code for Classifica- 
tion of Craniocerebral Injuries, Journal of the Michigan State Medical 

Society, Vol. 37, No. 10, pp. 909-912 (October, 1938). 
Smith, Bradford B., The Use of Punched Card Tabulating Equipment in 

Multiple Correlation Problems (U.S. Department of Agriculture, Bureau of 

Agricultural Economics, Washington, D.C., October, 1923). 24 pp. 
Tajdor, Erwin K., Use of a Single Card Column for Recording Variables with 

a Range of 30 or Fewer Units, Proceedings of the Research Forum, pp. 63-67 

(International Business Machines Corporation, New York, August, 1946). 
Toops, Herbert A., Annotated Bibliography of Tabulating and Recording 

Devices, Ohio College Association Bulletin, No. 122, pp. 2401-2410 (April 

13, 1940). 
Toops, Herbert A., Research Possibilities of Addends, Proceedings of the 

Research Forum, pp. 68-74 (International Business Machines Corporation, 

New York, August, 1946). 
Tucker, Ledyard R., The Centroid Method of Factor Analysis by Punched 

Cards, Bibliofilm Manuscript No. 1623 (University of Chicago Library, 

1941). 
Tucker, Ledyard R., A Note on the Computation of a Table of Intercorrela- 

tions, Psychometrika, Vol. 13, No. 4, pp. 245-250 (December, 1948). 
Tucker, Ledyard R., A Note on a Machine Method for the Quantification of 

Attributes, The Prediction of Personal Adjustment (Paul Horst, Editor), 

Bulletin No. 48, pp. 347-348 (Social Science Research Council, New York, 

1941). 



PUNCHED-CARD COMPUTING SYSTEMS 181 

Tucker, Ledyard R., Simplified Card Methods in Factor Analysis, Proceed- 
ings of the Research Forum, pp. 9-19 (International Business Machines Cor- 
poration, New York, August, 1946). 

Usilton, Lida J., A Mechanical System for Record Keeping of Morbidity, 
Treatment-Progress, and Control of Venereal Diseases, Venereal Disease 
Information, Vol. 21, No. 3, pp. 61-67 (March, 1940). 

Varossieau, W. W., Use of the U. D. C. in Selecting Data with Mechanical 
Appliances, Revue de la Documentation, Vol. 15, No. 2, pp. 41-46 (1948). 

Vickery, C. W., On Drawing a Random Sample from a Set of Punched Cards, 
Supplement to the Journal of the Royal Statistical Society, Vol. 6, No. 1, pp. 
62-66 (1939). 

Wallace, Henry A., and George W. Snedecor, Correlation and Machine 
Calculation, Official Publication of Iowa State College, Vol. 23, No. 35 
(January 28, 1925). 47 pp. 

Wangeman, C. P., An Experiment in the Recording of Surgical and Anes- 
thetic Data in Military Service. The Adaptation of Hollerith Punch 
Cards, Used at Second Army Maneuvers, 1940, Anesthesiology, Vol. 2, 
No. 2, pp. 179-185 (March, 1941). 

Warren, Richard, Hollerith Machine Computation of Least Square Trend 
Lines, Columbia University Statistical Bureau Document, No. 2 (1933). 
35 pp. 

Warren, Richard, and R. M. Mendenhall, The Mendenhall- Warren-Hollerith 
Correlation Method, Columbia University Statistical Bureau Document, 
No. 1 (1929). 34 pp. 



CHAPTER 10 
LARGE-SCALE DIGITAL COMPUTING SYSTEMS 

10-1. Introduction 

Four large-scale digital computing systems using electro- 
mechanical relays as primary elements in arithmetic, storage, 
transfer, and input-output are in operation: the Harvard 
Mark II, the Bell Telephone Laboratories' Models V and VI, 
and the IBM Pluggable Sequence Relay Calculator. The IBM 
Automatic Sequence Controlled Calculator, commonly called 
the Harvard Mark I, employs relays for switching and employs 
the IBM counter wheel for storage and for the arithmetical 
operations. Four large-scale electronic digital calculators are 
at present in operation: the ENIAC, the IBM Selective Sequence 
Electronic Calculator, the Harvard Mark III, and the BINAC. 

This chapter will concern itself with the discussion of some of 
these computing systems and others which are in design or under 
development. A brief outline of each computer will be given 
from the point of view of the types and numbers of components 
employed in the system. For more detail regarding the func- 
tions of those components, the reader is invited to examine the 
contents of Chaps. 3, 4, and 13 to 17. A list of major design and 
development projects on large-scale digital computers in this 
country is presented in Table 10-1, pages 214-215. 

10-2. Relay Computing Systems in Operation 

During the period 1944 to 1947, four relay computers were 
put into operation on scientific problems. These computers 
vary widely in complexity and character. In general terms, 
however, all of them are automatic sequence-controlled machines. 
The IBM Pluggable Sequence Relay Calculator, while normally 
not listed with large-scale digital machines, is included here for 
purposes of comparison with IBM equipment discussed in 
Chap. 9. 

The general characteristics of relay computers are reviewed 
in the literature. 21 

182 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 183 

10-2-1. Harvard Mark I Calculator. General. Design and 
construction of this calculator was begun in 1939 at Endicott, 
New York, and continued until August, 1944, when the machine 
was presented formally to Harvard University by T. J. Watson, 
President of the International Business Machines Corporation. 
Professor Howard H. Aiken of Harvard University, then Com- 
mander, USNR, on active duty, B. M. Durfee, F. E. Hamilton, 
and C. D. Lake were the coinventors of the machine. 1 

The relays, counters, cam contacts, typewriters, card feeds, 
and card punches employed in the calculator are all standard 
parts of the tabulator machinery as constructed by the Inter- 
national Business Machines Corporation. 1,3,17,;L8 

The external outline of the Mark I consists of a 51-foot panel, 
8 feet in height; two 6-foot panels extend back at right angles 
from the front panel. Along the 51-foot panel are mounted (1) 
the sequence control, which directs the programming by punched 
tape, (2) three interpolators, which are tape-fed units for selecting 
data required in the interpolation process, (3) functional counters 
for controlling the interpolation of functions, and for computing 
logarithmic and trigonometric functions and printing, (4) a 
multiplying-dividing unit for 23-digit accuracy, (5) 72 storage 
counters for the intermediate storage of results up to 23 digits 
capacity, used for arithmetic operations and accumulation, and 
(6) a storage unit with a capacity of 60 23-digit decimal numbers 
which can be introduced to the machine by manually setting dial 
switches. The mechanical drive for this calculator is a 4-horse- 
power motor ; the total assembly weighs approximately 5 tons. 

Input, Output, and Control. The machine operations are 
controlled by orders introduced on a coded 24-hole punched 
tape consisting of three columns of eight holes each. This 
input tape also controls the time required to add two numbers, 
since it is that time which is required for the control tape to 
move one step forward. The control tape can be advanced about 
200 units per minute, which establishes the fundamental time 
cycle of the machine as 0.3 second for the addition of two numbers. 
Three tape-driven interpolators control the routine of the inter- 
polation process in detail. One of the three tape-driven inter- 
polates may also be used to introduce numerical values to the 
calculator. 

The machine processes the results of its calculation through 



184 COMPUTING SYSTEMS 

the use of standard IBM equipment: two punch-card readers, a 
card punch, and two automatic typewriters. 

Arithmetic Operations. The Mark I Calculator can perform 
arithmetic operations on numbers up to 23 decimal digits, but 
instructions can be introduced into the machine for its operation 
to 12 or 46 significant figures. Addition and subtraction are 
performed with the set of 72 electromechanical adding-storage 
registers, which can store 23 digit numbers and can combine 
numbers which are generated by the machine in the course of 
its operation. Multiplication and division are performed in 
the multiplication-division unit, and the process followed is that 
of building up a small table of the first nine multiples of the 
multiplicand or divisor and using this table in an appropriate 
manner. Multiplication requires approximately 19 cycles, or 
approximately 5.7 seconds; division requires about 51 cycles, or 
15.3 seconds. 

Function tables of log x, 10*, and sin x can be constructed 
using an array of relay elements. Logarithmic tables are 
computed by using four factors of the form 

1 + h • 10-* where k = 0(1)3 and h = 0(1)9 

and the power series for log (1 + x). [The notation a(b)c is used 
in describing the extent and mesh of the arguments appearing 
in a table of functions. The first number a in the notation 
and the last number c are the first and last arguments in the table ; 
the number b in parentheses is the difference between the argu- 
ments of successive entries in the table. Thus, for a standard 
set of five-place logarithms with four-place arguments, the 
notation is 1.000(0.001)9.999.] This process reversed gives 10*. 
A power-series method is used to compute sin x or cos x for values 
of the argument less than 45°; other values of trigonometric 
functions are derived from sin x or cos x. Calculation of loga- 
rithmic and simple trigonometric functions requires about 200 
cycles or 1 minute. 

Storage and Transfer. Sixty 23-digit decimal numbers can be 
stored in registers; numbers are set into the registers manually. 
The 72 registers used in the arithmetic operations provide inter- 
mediate storage. Numbers are transferred from one part of the 
machine to another by timed electrical pulses of 50 volts ampli- 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 185 

tude on a single bus. It is not possible, therefore, to transfer 
simultaneously two or more numbers. 

10-2-2. Harvard Mark II Calculator. General. This cal- 
culator was designed and constructed by Harvard University 6 
for the Naval Proving Ground, Dahlgren, Va. The design 
was initiated in early 1945; construction was completed in 1947. 
The machine is now in operation at the Naval Proving Ground. 

The Mark II uses electromechanical relays for the internal 
storage of numbers, for the transfer of numbers, for performing 
the basic arithmetical operations, and for sequence control of 
these processes. The machine handles 10-digit decimal numbers 
and uses the floating decimal point. The addition time for two 
10-digit figures is 125 milliseconds and for two 10-digit multi- 
plications, about 250 milliseconds. About 100 10-digit decimal 
numbers can be stored in the machine; additional internal or 
intermediate storage may be supplemented by punching data 
on tapes for introduction to the machine at a later time. The 
machine registers its results on punched tape or through the use 
of page printers in the case of tabular material. It is interesting 
to note that the Mark II can be operated as a whole on one prob- 
lem or can be split into two parts and each operated separately. 

The Mark II contains about 13,000 electromechanical relays. 
These relays, specially constructed for the machine, are six-pole 
high-speed units with double-throw contacts. They operate 
on 100 volts direct current and take about 6 watts when con- 
tinuously energized. The operating time of these relays is 
relatively short, varying from 6 to 10 milliseconds. One-third 
of the relays are provided with mechanical locking devices which 
hold them in position without the relays being energized by a 
steady current. These latches or locks are used in the internal 
memory of the machine to reduce power requirements and, 
in the case of a power failure, to provide a nonvolatile storage 
system. 

Input and Output. The calculator has 12 input mechanisms for 
introducing commands and numbers to the machine. All these 
mechanisms employ a punched-paper-tape input. Commands are 
introduced to the machine on one group of four input tapes. 
Another group of four input tapes introduces numbers to the 
machine in any desired preassigned order. The group of four 
input tapes is used for supplying the calculator with coded tables 



186 COMPUTING SYSTEMS 

of functions. Eight hundred functional values can be introduced 
on these punched tapes. 

The calculator records the final results of a computation 
through the use of four page printers or automatic typewriters. 
Each of these printers can record one complete number (25 char- 
acters) in about 4 seconds. If the results are to be used again 
by the machine itself, they may also be automatically punched 
into paper tape. Four tape punches are available for this pur- 
pose and can be made to operate at a maximum rate of about two 
quantities in every 3 seconds. 

Number System and Storage. A number N is represented by 
the calculator in terms of another number p and an integer j 
where 

N = p X 10'', 1 < p <10 and -15 < j < +15 

In order to represent p, four relays are required for each of the 10 
decimal columns and one relay for the algebraic sign; j is repre- 
sented in the binary system, which requires four relays. A 
fifth relay for j is used to denote algebraic sign. A storage 
register, therefore, requires 46 relays with 16 additional relays 
for input and output of information to the register. Since there 
are 100 such storage units in the machine, there are 6,200 relays 
in the complete storage-register system. The output time for 
any number in any register is approximately 33 milliseconds. 

Six algebraic and transcendental functions are stored per- 
manently within the machine; the reciprocal, the reciprocal 
square root, the logarithm, the exponential, the cosine, and the 
arc tangent. These functions can be calculated for any argument 
to eight or nine significant figures as long as the argument and 
the function are within the digit capacity of the machine. These 
functions may be computed to this degree of precision within 
from 5 to 12 seconds. 

Arithmetic Operations. Because of this method of representing 
numbers in the machine, the process of addition of two numbers 
has to be accomplished in two steps: The first step is a comparison 
of the exponents j by the machine, and the difference between 
the smaller and larger exponent is obtained. The number char- 
acterized by the smaller exponent is then shifted by the amount 
of the difference between its j value and that of the larger number. 
After an addition of two numbers has been performed, the sum 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 187 

may have to be shifted as it is placed into storage. Negative 
numbers are represented by complements on 9's; these are 
obtained in the adder itself, since the storage units do not use 
the complementary system. Addition requires about 200 milli- 
seconds; this time is divided approximately into an input time 
of 33 milliseconds, an operation time of 125 milliseconds, and 
output time for the sum of about 33 milliseconds. The Mark II 
Calculator contains two units for addition. 

Multiplication in the Mark II is as described by R. V. D. 
Campbell 6 : 

"In multiplying, the calculator first obtains five multiples 
of one of the factors, called the multiplicand (mc). Of these 
multiples, 1 X (mc), 2 X (mc), 4 X (mc), and 5 X (mc) are 
obtained by circuits which multiply directly, but 3 X (mc) 
is obtained by addition. Each digit of the multiplier, working 
from right to left, selects the proper multiple of the multiplicand, 
which multiple, after the appropriate shift, is either added to 
or subtracted from the sum of the previously chosen multiples. 
Note that, because only five multiples are used instead of nine, 
it is necessary to be able to subtract them as well as to add them. 
The subtraction is carried out by using complements on nine. 
The exponent of the product is equal either to the sum of the 
exponents of the factors or to that sum increased by one. Mul- 
tiplication requires 0.7 second, regardless of the magnitudes of 
the factors or of how many non-zero digits they contain. There 
are four multipliers in the calculator." 

10-2-3. Bell Telephone Laboratories' Relay Computing Sys- 
tems. General. The Bell Telephone Laboratories began the 
design of their computing systems in 1938, using the apparatus 
and circuit design techniques used generally in dial-telephone 
systems. Six types of computers, recently designated by Model 
numbers I to VI, have been built. Two Model V computers 
have been built and one of each of the others, making a total of 
seven. All of these except the Model I are in operation. Models 
V and VI are in the large-scale computing machine class, and 
brief descriptions of them are given below. 

Of the earlier models, Model I was completed in 1940 and was 
demonstrated before the Mathematical Society meeting of that 
year at Dartmouth. The demonstration consisted of introducing 
complex numbers into the computer by telegraph to New York 



188 COMPUTING SYSTEMS 

and receiving the answers through a telegraph printer at the 
meeting. This machine was recently dismantled when the more 
versatile Model VI was put in operation. 

Model II (1943) and Model IV (1945) are in operation at the 
Naval Research Laboratory in Washington, D.C. Model III 
(1944) is in operation at the Laboratory of the Army Field 
Forces Board at Fort Bliss, Texas. 

Model V. In 1944 the development of an all-purpose auto- 
matic computing system 2,25,28 was started. This computer con- 
tains more than 9,000 relays and about 50 pieces of teletype 
equipment, weighs about 10 tons (twice the weight of the Mark I), 
and covers about 1,000 square feet of floor space. 

Two models of the BTL Model V machine have been built. 
Each of these models consists of two computers with the necessary 
equipment. Each computer contains all the components needed 
for dealing with a computational problem with the exception 
of tape-reading devices which are used in common between the 
two computers. Each contains 15 storage registers; eight sign 
registers; a calculator for the basic arithmetic operations and 
square-root extraction; a routine control for directing the execu- 
tion of instructions to the computer; a BTL register, for block- 
forming, or number-grouping, trigonometric, and logarithmic 
operations; a table register and table control for receiving and 
storing numbers read from table tapes; a problem-registering 
problem control for receiving and storing numbers from sections 
of the problem tape; a printer-registering printer control for 
receiving numbers from various parts of the machine and direct- 
ing the process of printing and perforating; a discriminator, 
which selects one out of several possible methods of computation 
and directs the machine to perform one; a recorder table, which 
contains a printing reperforator, a tape transmitter, and a 
distributor. 

For problems of medium complexity, usually one computer of 
this system is used for carrying out the computation. For large 
problems, the two computers are associated by the problem 
programming, and both are used simultaneously during the 
computation. 

Input and Output. The input, similar to that for the other 
digital machines, consists in introducing both numbers and 
orders to the machine on a perforated paper tape; the input speed 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 189 

is approximately that of the computing speed of the machine. 
About 2 seconds are required to read and transfer into the 
machine a seven-digit number or a command of average length. 
Addition can be carried out in 0.3 second, multiplication in about 
1 second. Input speed of the numbers, therefore, is slightly 
slower than the unit time of operations for the machine. 

The output of the machine is performed by standard teletype 
devices called reperforators, for punching numerical data on tape, 
or page printers, for printing on paper. 

Number System. The essential difference between the Bell 
Telephone Laboratories system and the other electromechanical 
calculators is number representation in the biquinary system. 
In this system, a decimal digit is replaced by two digits; one of 
these digits is a quinary digit which has one of the values from 
to 4; the other is a binary digit which has one of the two values 
or 5 in such a way that the sum of the two values is equal to 
the value of the decimal digit which it represents. Since there 
are five binary relays and two quinary relays in this system, the 
10 decimal digits, to 9, require the availability of seven relays 
for their registration, although actually only two relays are 
activated at any one time. This represents a saving of three 
relays over a machine operating on the decimal system where 
10 relays must be available. Besides the seven relays, two 
others are required for representing the sign of the number; the 
negative numbers are not represented by complements but by 
the number itself preceded by a minus sign. 

Like the Harvard Mark II Calculator, the BTL machine uses 
the floating decimal point. Each number is transformed into a 
decimal number with seven significant digits multiplied by the 
appropriate power of 10. The first decimal digit transferred 
lies to the right of the decimal point, and the exponent of 10 
provides that number by which the decimal is to be multiplied. 
In this machine the exponents include the range —19 to +19. 
These exponents are represented by a positive or negative sign, 
a units digit and a 10's digit. The 10's digit, 0, and 1, and the 
sign, + or — , are each stored on one of two relays. The units 
digits, to 9, are stored in the biquinary system on two of seven 
relays. 

Arithmetic Operations. The process of addition is carried out 
using three groups of relays each of which can store a 10-digit 



190 COMPUTING SYSTEMS 

number in the biquinary system. These groups are so wired 
that if numbers are introduced to the first two groups, the sum 
automatically appears in group three. In subtraction the 
process is conducted by converting the number with the larger 
exponent into its complement with respect to 9,999,999. The 
other number is introduced to the second group of relays shifted 
as required with respect to the exponents. The result, which 
appears in the third group of relays, is the complement of the 
desired number, and the machine must convert it back into the 
number required. Multiplication is performed by repeated 
addition, and division is performed by repeated subtraction. 
The extraction of square roots is also performed by repeated 
subtraction. 

The estimated speeds of several arithmetic operations for 
seven-digit numbers are given below: 

Operation Time, seconds 

Addition or subtraction 0.3 

Multiplication 1.0 

Division 2.2 

Square root 4.3 

Number transfer from one register to another register . 07 

These estimates include the time required to introduce a seven- 
digit number into the calculator and place a seven-digit number 
(result) in a given register. 

Storage. The machine can store 30 seven-digit decimal 
numbers. A unit of 62 relays, required to store one seven-digit 
decimal number, is called a register; the BTL machine has 44 
such registers, 30 for number storage and 14 for performing other 
functions. Each computer is also equipped with a register 
called a BTL register which serves as a storage medium and 
performs special operations such as those of forming block 
numbers and computing trigonometric and logarithmic functions. 
In this sense, the BTL register is closely related to the permanent 
function tables in the machine. 

These functions, which are available to the machine in storage, 
include sin x, cos x, and tan -1 x. The machine for the Ordnance 
Department of the Army contains, in addition to these functions, 
log x and 10 x . 

The setup time has been reduced by provision of more problem 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 191 

setup positions than can be used at one time. In the Aberdeen 
installation there are two computers, but there are four problem 
positions. With this arrangement two problems can be set up 
on the two idle positions while the computing equipment is 
actively using the other two. As soon as either one of the two 
problems has been completed, one of the stand-by problems is 
automatically connected in. This automatic feature is one of 
the reasons why the Model V computers can be operated on an 
unattended basis. 

Model VI. The Bell Telephone Laboratories has recently 
placed in operation in its Murray Hill, N.J., building a new 
digital system called the Model VI. Remote-control stations 
are located in the quarters of the computing staffs, and the com- 
puting equipment itself is located in another part of the building. 
A switchboard cable of the type used in telephone central offices 
is used to provide for the exchange of information between the 
two locations. The machine contains about 4,300 relays, 86 
cold-cathode tubes, and other miscellaneous apparatus. This 
machine is a little less flexible than the Model V but is simpler 
in design detail. Each number is introduced into the machine 
as a three-, six- or 10-digit decimal number times a power of ten 
ranging from —19 to +19. The floating decimal point appears 
to the right of the most significant digit. 

10-2-4. The IBM Pluggable Sequence Relay Calculator. 
General. The IBM Corporation developed two models of this 
relay calculator 7 during the war; they were used at the Aberdeen 
Proving Ground during 1944 to 1945. In 1945, three additional 
machines, with improvements over the first design, were built. 
One of these machines is now at the Naval Proving Ground, 
Dahlgren, Va. The other two machines are at the Watson 
Scientific Computing Laboratory, Columbia University, in the 
City of New York. 

Input and Output. This relay calculator operates on numbers 
read from punched cards, undertakes a sequence of calculations 
by relay networks, and punches the result of these calculations 
on cards. The basic operations of the machine are synchronized 
with the rotations of a shaft in the card-reading unit. This 
shaft rotates at a speed of 100 revolutions per minute. The card 
feeds and punches are synchronized to operate within the time of 
a single rotation of a shaft. During one shaft rotation, a timing 



192 COMPUTING SYSTEMS 

circuit generates 48 impulses spaced at 0.0125 second; these 
impulses operate the relays. Another timing circuit generates 14 
pulses during a shaft rotation which is used to actuate the reading 
and punching mechanisms. With this system one impulse for 
the card-actuating mechanism is equivalent to 3% relay impulses. 

Arithmetic Operations. The operations of addition and sub- 
traction require two impulses and three relay groups, one for 
the augend, one for the addend, and one for the sum. If suc- 
cessive additions or subtractions involve the same relay, four- 
group relay impulses are required. Multiplication is performed 
in a manner similar to that used in the IBM multiplying punch, 
Type 601 (see Chap. 9). Division is performed by subtracting 
the divisor or 5 times the divisor from the dividend. In division, 
the dividend is shifted to the left after the completion of each 
subtraction operation in one position. The process of subtrac- 
tion consists in adding the complement of the number to be 
subtracted. The capacity of this calculator for division allows 
for six figures, 12 digits in the dividend, six in the divisor, and 10 
in the quotient; the time for division is approximately 0.2 second 
per quotient digit. The extraction of a square root is performed 
by successive additions of odd numbers to the complement of the 
original number. In the square-root process the capacity of the 
calculator is 12 digits in the original number and six digits in the 
square root. 

Storage and Control. The machine has a total of 36 storage and 
computing registers. The normal capacity of the machine is 
six-digit decimals, although higher precision than this may be 
obtained with sequence operation. In conjunction with the 
punched cards the plugboard facilities connected with the 
machine allow operations to be performed in parallel rather than 
the straight sequence operation resulting from punched-tape 
input. In comparison with the standard IBM Calculating 
Punch (see Chap. 9 for greater detail), this calculator has a 
higher operating speed. It is approximately ten times as fast 
as the calculating punch. 

The sequence-control operations can be initiated by a pluggable 
sequence and also by the punches on the cards. Actually the 
pluggable sequence on the control panels starts the operations. 
In this system, there are 96 sequence relays connected to the 
plugboard so that the impulses through these relays can be used 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 193 

for control purposes. The relays operate in succession; for a 
complete sequence of 96 relays to be operated, the time required 
is two shaft rotations of the basic drive, or 20 milliseconds. In 
cases where double operations occur, such as multiplication and 
addition, whose orders are required to be read from the card and 
whose results are to be printed on the card, the machine may be 
controlled to operate at 50 cards per minute through the reading 
and punching units. 

10-3. Electronic Computing Systems in Operation 

In order that more complicated physical and mathematical 
problems may be solved in an economical time interval, the unit 
operation time for automatic computing systems 15,20 must be 
decreased wherever possible. By using electronic vacuum 
tubes for switching, storage, and for arithmetic operations, a 
saving in time in these functions of several orders of magnitude is 
effected over similar devices using electromechanical relays. 

It is probable that, for some time to come, input and output 
mechanisms using data punched on cards or paper tape or 
registered as dots on film or as marks on a magnetic medium 
will have to be employed. Although such physical systems 
operate at inherently slower speeds than electron vacuum tubes, 
their use is not as stringent a time-limiting element as would be 
thought at first. For solving highly complicated mathematical 
problems, which may represent in reality the interaction of 
many physical parameters, it is desirable that the computer 
retain as many intermediate results as possible during the com- 
putation. This can be accomplished in an electronic computer 
if mechanical devices are avoided for use as intermediate storage 
mediums. For example, there would be no reason for employing 
high-speed electronic arithmetic circuits if numbers and commands 
are presented to or abstracted from them by devices whose unit 
operation time is much slower than that of the electronic vacuum 
tube. If mechanical devices are limited in their application 
only to the initial input numbers and commands and to the final 
registration of completed results and if switching, gating, and 
arithmetic operations are conducted at speeds which are of the 
same order as electronic-vacuum-tube operation, the computer 
may carry out thousands of calculations within itself before 
arriving at a result worthy of physical representation outside 



194 COMPUTING SYSTEMS 

the machine. In such a calculation the machine may perform 
a large number of arithmetic operations; it may construct func- 
tion tables or generate new numbers and orders in sequence; it 
may also store intermediate results and draw upon them when 
required. Such a procedure would warrant the complicated 
electronic circuitry which speed and the consequent reduction of 
calculating time demand. Although with present input mecha- 
nisms it requires a considerable length of time to code, instructions 
and numbers for input to the machine, the total time for 
these operations may be made feasible if the machine can carrj/ 
out calculations a thousand times as fast within itself; in the 
same way a mechanical output, either a punch or a printer, 
must not hold up the internal functions of the calculator. For 
each number punched or printed, there must be at least 1,000 
operations performed electronically within the machine. If the 
machine presents data to the output faster than this, an inter- 
mediate-speed storage mechanism is then required for protecting 
the computer against slow-output intrusions upon its high-speed 
capabilities. The intermediate storage effectively changes the 
time scale from that within the computer to that outside of it. 

10-3-1. The ENIAC (Electronic Numerical Integrator and 
Computer). General. This computer was developed at the 
Moore School of Electrical Engineering, University of Pennsyl- 
vania. 12 ' 16 ' 23 It was first demonstrated to the public in February, 
1946, and was transferred in 1947 to the Ballistic Research 
Laboratories, Aberdeen Proving Ground. 

The ENIAC is used, essentially, as a general-purpose com- 
puter. As an example which can be regarded as an illustration 
of the upper limit of its capabilities, it can be used to solve the 
system of five simultaneous hyperbolic partial differential equa- 
tions describing flow around a body of revolution. For each case 
{i.e., each different mach number and set of shape parameters) 
this problem takes the machine about an hour. 

The ENIAC is the first large-scale machine to make use of 
electronic circuits for general operations except for input, out- 
put, and certain switching functions. There are 40 separate 
panels arranged in a U configuration with a total of approxi- 
mately 1,500 electromechanical relays and 18,000 vacuum tubes. 

As was indicated by the footnote on page 35, the program- 
ming philosophy of the ENIAC has been modified since it was 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 195 

first installed at the Ballistic Research Laboratories. The 
machine itself has not been modified except by the addition of a 
converter. This unit decodes pairs of decimal digits and, for 
each pair of decimal digits, energizes a unique output cable. 
There are 100 of these cables. 

A signal on any one of the 100 conducting cables which can be 
energized by the converter initiates a particular operation. The 
portable function tables which were formerly used to store tables 
of functions are now used to store sequencing instructions in 
addition to numerical data, tables of functions, constants, etc. 
Each instruction consists of a pair of decimal digits which are 
decoded by the converter unit. The pairs are read one at a time 
in accordance with the program sequence. 

Each output cable from the converter unit initiates an oper- 
ation which has been determined by the arrangement of the 
plug connections, and the settings of the switches are called the 
background wiring. The background wiring is changed only if 
it is desired to redesign the coding. 

To instruct the machine, then, it is necessary only to set the 
switches on the portable function tables. The operations to be 
performed and the sequence in which they are to be performed 
are specified by the switch settings. 

Input and Output. The numerical data to be introduced to the 
machine are first punched on standard IBM cards; these cards 
contain 80 columns, each column with 12 punched designations 
(see Chap. 9 for details on IBM equipment). Ten of the 
punched designations, to 9, are used to represent one decimal 
digit. Positions 11 and 12 may be used for representing the 
sign of a digit in one column or a group of digits in blocks of 
adjacent columns. The 11 and 12 punch holes may then be 
used for representing a positive or negative sign for each of 
these numbers; it is seen that 16 such punch positions must 
therefore be available. The numbers represented by card 
punches are then read and stored in relays in the constant trans- 
mitter, which makes these numbers available to the various 
operating components of the machine when required. In a 
similar way, the results which have been calculated by the 
ENIAC are punched on cards using an IBM card punch; the 
data from these cards can then be printed automatically through 
the use of an IBM tabulator. 



19G COMPUTING SYSTEMS 

Arithmetic Operations. All the units which perform numerical 
operations — addition, subtraction, multiplication, division, and 
extraction of square root, together with the capacity to obtain 
numbers from the machine's storage system — operate by using 
electronic vacuum tubes on the flip-flop principle. For addition 
and subtraction, the machine makes use of 20 accumulators. 
Each one of these accumulators can hold a signed 10-digit 
decimal number. An accumulator is capable of receiving a 
number transmitted to it and adding this number to the number 
which it already contains. Also, on one output channel, the 
accumulator can transmit the number stored at any given time. 
On another output line the accumulator can transmit the com- 
plement of the number it stores. As described in Chap. 3, the 
10-stage (decade) ring counter is designed to count pulses in a 
one-way progression around the ring. Subtraction of two 
numbers, therefore, is most effectively performed with decade 
ring counters through the use of complements. 

The numerical circuits for the multiplier make use of four 
accumulators, or up to six if products with 10 to 20 digits are 
required or if the best 10-digit accuracy is required. The units 
digits of the partial product of the entire multiplicand by one 
digit of the multiplier are transmitted to one accumulator while 
the 10's digits are transmitted to another accumulator. For 
example, to multiply 567 X 234 the machine first multiplies 
567 X 4. A novel method, which is much faster than the normal 
multiplication by successive additions, is employed in accumu- 
lating this partial product. 

The units digit of the product of 7 X 4 or 8 is transmitted to 
one accumulator and the 10's digit or 2 to another. Similarly 
the units digit of the product of 6 X 4 or 4 goes to the units 
accumulator shifted one digit to the left and the 10's digit, 
again 2, goes to the 10's accumulator also shifted. As a result 
of the multiplication of 567 X 4, therefore, 048 is stored in one 
accumulator and 222 in the other. For 567 X 3, 581 is stored 
in one accumulator, 112 in the other (shifted with respect to the 
numbers already accumulated). Two multiplication tables are 
employed; one produces the 10's digits of the required product, 
and the other the units digits. When the digits of the multiplier 
have been used up, the 10's and units digits of the partial products 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 197 

are combined in one of these accumulators to obtain the total 
product. 

The divider is one panel of equipment which controls neighbor- 
ing accumulators so as to perform division or extract a square root. 
One accumulator stores the numerator (radicand), one stores the 
denominator (or twice the square root), one stores the quotient, 
and finally, a fourth accumulator is used for shifting operation. 
Division is accomplished by a method similar to that used in some 
desk machines. The denominator is subtracted from the 
numerator repeatedly until an overdraft appears. The remainder 
is then shifted one place, and the denominator is added until this 
overdraft is wiped out. The quotient accumulator counts the 
number of cycles in each place. The process of extracting the 
square root of a number is carried out by a combination of suc- 
cessive subtraction and the addition of odd numbers, analogous 
to the process used in some desk calculating machines. 

Storage. The storage system is composed of units which store 
numerical tabular data in electronic circuits; it also contains elec- 
tromechanical switches for storing instructions for a specific 
problem. There are two additional units, one for initiating 
computation, for clearing the machine, and for testing; the other 
is a cycling unit which supplies fundamental signals to the 
machine and synchronizes their operation. The fundamental 
unit of time in the ENIAC is Kooo second; during this period a 
pattern or train of pulses is emitted. One cycle is referred to 
in the ENIAC as an addition time. The constant transmitter, 
for example, requires one addition time to emit a signed five- or 
10-digit number. 

The pulse width is 2 microseconds and the machine is operated 
at about 100 kilocycles per second. 

The machine, as now used, has two internal storage systems 
from which it can withdraw stored numbers called for by a par- 
ticular program of operations. The basic difference between 
them is that data can be stored in one of these by the machine 
itself; data can be stored in the other only by the operator. As 
indicated under the general description of the machine, the port- 
able function tables (which are simply resistor matrices) are 
used for the storage of numerical data as well as for sequencing 
instructions. They comprise the second type of internal storage, 



198 COMPUTING SYSTEMS 

in which data can be inserted only by the operator. The other 
storage system consists of accumulators. 

10-3-2. The IBM Selective Sequence Electronic Calculator. 

General. This calculator, developed by the International Busi- 
ness Machines Corporation, 19 was put into operation in January, 
1948. It combines electronic-vacuum-tube techniques and 
electromechanical relays. Vacuum-tube circuits are employed 
for arithmetic processes and for high-speed storage, switching, 
and other control orders. Relays and punch tape form secondary 
storage units. 

Input and Output. The instructions are introduced to the 
machine either from the standard 80-column IBM cards or from 
continuous card-stock tapes. The keyboard punch for these 
tapes may be operated manually by keyboard or may be operated 
automatically by punched cards. The punch converts a decimal 
number into its binary equivalent, which is then handled by the 
calculator. The card-stock tape is approximately the width 
of the long dimension of the standard IBM card, allowing 80 
perforations per line across the tape. This provides ample space 
for representing a 19-decimal-digit number together with alge- 
braic sign across one line of the tape. The results of machine 
calculations can be punched on standard IBM cards or can be 
printed in tabular form. If it is required, both methods may be 
used. The digits can be punched on standard IBM cards at 
the rate of 1,600 digits per minute and can be printed in tabular 
form at the rate of 2,400 digits per minute. 

The operating instructions for the machine are read from 
punched tapes, and groups of commands are delivered at the 
rate of 50 command groups per second to the machine. Since 
these commands must be synchronized with the accepted arith- 
metic transfer and storage operations, a 20-millisecond delivery 
rate appears entirely feasible in view of the electromechanical 
devices whose operating characteristics must be tied in with the 
other machine components. The array of orders and numbers 
on input tapes does not appear to differ greatly in procedure 
from other machines which employ punched-tape feeds. 

Arithmetic Operations. Addition, subtraction, multiplication, 
and division are performed by electronic-vacuum-tube counting 
circuits. The representation of the numbers in the machine is 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 199 

not specified, but it is stated that this calculator contains more 
than 100 electronic counting units which can be combined for 
arithmetic operation to add or subtract two 19-digits in less than 
10 milliseconds. The electronic counting units will multiply 
two 14-digit numbers to give a 28-digit product in 20 milli- 
seconds. For division these units may be combined to allow 
the division of two 14-digit numbers to give a 14-digit quotient 
in about 30 milliseconds. Provision for handling decimal points 
and indicating the algebraic sign of each number is made. 

Storage. The internal storage capacity of the machine totals 
400,000 digits and is divided among electronic, electromechanical, 
and punched-tape units. In electronic storage there are eight 
units, each of which can store one 19-digit decimal number with 
algebraic sign. About 3,000 digits (150 numbers) can be stored 
in the electromechanical relay units. Punched-tape units can 
store up to 396,800 digits (20,000 numbers). 

The punched-tape storage system comprises three units, 
each with a punching unit for the card-stock tapes and each with 
10 reading stations. The total of 30 reading stations may be 
combined in a number of ways to permit their use on individual 
or specific problems. On a single line of punching on the tape, 
it is possible to store 19 digits and indicate the algebraic sign of 
the number. 

For the table-consulting function, the machine is provided with 
a group of 36 reading stations additional to those identified 
with the storage system. These reading units are used to 
locate specific information on any of the 36 punched tapes flowing 
through each reader. Since each of the 36 reading units may 
have a tape with capacity for 150 19-digit numbers, the total 
capacity for the tabular system is 5,400 numbers, or 102,600 
digits. These units may be combined so that a maximum of 
38 digits with algebraic sign may be made out. The access 
time for this table or for searching the 36 tape units is about 
3 seconds. Approximately 160 milliseconds is the time for the 
tabular unit to find a particular value in an eight-place sine table. 
In the arrangement of the tabular information, on punched 
tapes, the values of the argument of the function with variable 
intervals may be used, which in some cases may allow the use 
of fewer tabular values. In addition to the punched-tape 



200 COMPUTING SYSTEMS 

storage, intermediate results, which may not be required for 
reintroduction into the machine for a considerable length of time, 
may be stored on IBM cards. 

10-3-3. Harvard Mark III Calculator. At the time of this 
writing the Mark III is in operation at the Harvard Computation 
Laboratory where it was built for the Navy Bureau of Ordnance 
for eventual installation at the Dahlgren Proving Ground. This 
machine was made available for inspection by visitors during the 
Symposium on Large-Scale Digital Calculating Machinery which 
was held at Harvard on September 13 to 16, 1949, under the 
joint sponsorship of the Bureau of Ordnance and Harvard 
University. 

The plans for the Mark III computer were discussed by Dr. 
Aiken at the 1949 IRE National Convention. The Mark III 
was designed to provide greater speed and reliability, more 
flexible storage facilities, and greater ease of preparation of 
input data than were found in the earlier Harvard Computers. 1 
It was planned that the Mark III would use a magnetic drum 
for storing 4,000 orders and that 30 types of basic circuits would 
be incorporated into plug-in units, with color coding, for ease of 
maintenance. 2 

A preliminary description of this machine by B. L. Moore will 
be found in Proceedings of a Second Symposium on Large-scale 
Digital Computing Machinery (Harvard University Press), which 
is in preparation at the time of this writing. 

10-4. Electronic Computing Systems in Design or under Devel- 
opment 

The following electronic computing systems, in various stages 
of development, advanced design, or preliminary consideration, 
have been outlined here to indicate the trend in component use 
and the integration of those components into automatic sequence- 
controlled computers. 

10-4-1. The EDVAC (Electronic Discrete Variable Com- 
puter). General. This computer is the second high-speed 
electronic digital machine which has been constructed at the 
Moore School of Electrical Engineering at the University of 
Pennsylvania. 24 Although the entire calculator has not as yet 
been operated as a unit, many of the components have already 
been tested successfully, and the scheme of operation has been 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 201 

assured. From point of size, the EDVAC occupies much less 
space than the ENIAC and is constructed with panels 7 feet high 
occupying some 140 square feet of floor space. The number of 
electronic vacuum tubes in the EDVAC is only about 3,500, 
roughly one-sixth the number employed in the ENIAC. Out- 
side of the method for programming, in which a serial rather 
than a series-parallel arrangement appears to be the goal, the 
principal difference lies in the use of acoustic delay lines for 
internal storage of numbers. The EDVAC is designed to handle 
44-digit binary numbers. 

At the time of this writing, the EDVAC has been accepted by 
the Ordnance Department, Department of the Army, and has 
been installed at the Ballistic Research Laboratories at Aber- 
deen, Md. where it is now undergoing extensive tests. 

Input and Output System. As originally designed, binary 
digits (representing numbers and commands in four-address code) 
were intended to be read into and out of the machine on magnet- 
ized wire. This method was found to be unreliable, and teletype 
tape will be used as an interim input-output system. 

Storage. The EDVAC uses acoustic delay lines for high- 
speed storage; 126 mercury delay lines will provide storage 
capacity for 1,024 44-digit binary numbers. A single delay 
line will circulate eight numbers each of 44 binary digits. The 
pulse width is 0.3 microseconds (as compared with 2 micro- 
seconds in the ENIAC), and the repetition rate is 1 megacycle 
per second. 

Arithmetic Unit. Addition, multiplication, and division are 
built into the computer; these operations are based on the serial 
handling of digits. Total addition time, including transfer of 
numbers, is about 1 millisecond; multiplication and division 
require about 3 milliseconds. Double precision multiplication 
and division are built into the arithmetic units, enabling the pro- 
grammer, at his option, to perform operations yielding 88-digit 
results. A higher degree of precision can thus be achieved if 
desired. Checking is provided for by the simultaneous use of 
two algebraic units with inputs to a comparison unit. 

10-4-2. Electronic Computing Instrument (Institute for Ad- 
vanced Study). General. Preliminary information concerning 
the design of an electronic digital computer has been recorded 
by Burks, Goldstine, and von Neumann. 6 The planning and 



202 COMPUTING SYSTEMS 

coding of problems for such a computer have been treated by 
Goldstine and von Neumann. 13,14 It is not intended here to 
deal with the details of the electronic circuits or of the functioning 
of any given unit within this machine except as required to 
understand the principles of operation and the role to be played 
by each unit within the machine. 

In the preliminary exposition describing the design of an 
electronic computer the authors indicated that a fully automatic 
general-purpose computing machine requires a storage system, a 
control unit, arithmetic unit, and input and output units. At 
the time of this writing, a storage capacity of 1,024 numbers, 
each of 40 binary digits, has been decided upon as an interim 
solution to this particular design question, although 4,000 or 
more would be desirable. The computations which have been 
considered in order to arrive at this figure include the solution of 
(1) total differential equations, (2) partial differential equations 
of the elliptic, parabolic, or hyperbolic types, and (3) problems 
involving solution by iterative processes including systems of 
linear algebraic equations. 

Input and Output. The numbers and orders to the computer 
will first be coded on punched paper tape using modified teletype 
equipment. These coded numbers will then be transferred by a 
conversion unit to marks on a magnetic wire. Numbers and 
orders in this form are then introduced to the computer elements. 
For output, the numbers registered on the magnetic wire by the 
computer are transferred to punched paper tape. Some of the 
units for handling the punched paper tape in this input-output 
system were developed at the National Bureau of Standards. 

Storage. The proposed machine will have an electrostatic 
storage system consisting of a bank of Williams tubes (of the 
type described in Chap. 14 on pages 366-369). Test operations 
with cathode-ray tubes as electrostatic storage elements have 
already been conducted using the Williams technique. 

The Arithmetic Unit. The arithmetic units will operate in a 
parallel manner on binary numbers. At the time of this writing, 
an accumulator with capacity for 11 binary digits has already 
been tested, and the 40-stage arithmetic unit is now virtually 
complete and has already undergone extensive tests. 

The bank of Williams tubes operating in parallel as a storage 
system should make possible a very high total arithmetic speed; 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 203 

total addition time is expected to be of the order of 90 micro- 
seconds for two 40-digit binary numbers (including the time 
required to withdraw three numbers from storage and to return 
one), total multiplication time approximately 300 to 400 
microseconds. 

10-4-3. The UNIVAC (Eckert-Mauchly Computer Corpora- 
tion). General. The UNIVAC is a high-speed electronic com- 
puter 8 making use of acoustic-delay-line and magnetic-tape 
storage. At the time of this writing, practically all production 
drawings have been completed, and the first of six UNIVAC 
systems is now under construction. All portions of this system 
which are not similar to the BINAC (the computer built for 
Northrop Aircraft, Inc., by the Eckert-Mauchly Computer 
Corp.) have been laboratory tested in one way or another. 

The basic pulse rate for all internal circuits in the UNIVAC 
is 2.25 million pulses per second. The computer handles char- 
acter groups of 12 characters each; one such group is usually 
made up of 11 decimal digits and algebraic sign, but alphabetic 
letters and punctuation marks are also acceptable characters. 
The high-speed storage capacity is 1,000 character groups. Mag- 
netic tapes store an additional 10 6 groups under computer control. 

Input and Output. The instructions, and numerical or other 
data, are introduced to the computer by signals coded on metallic 
eight-channel magnetic tapes. More than 10 6 characters 
(decimal digits, letters, or other typewriter symbols) may be 
recorded in binary form on one reel of such tape 1,200 feet long, 
J<2 inch wide, and less than 2 mils thick. The reels are approxi- 
mately 8 inches in diameter. A keyboard resembling a type- 
writer keyboard is used as the input to the apparatus which 
records the coded signals on the magnetic tape. 

Normally, provision is made for computer connections to 10 
servo units, each of which controls one reel of tape. Altogether, 
more than 10 7 decimal digits (or other characters) can therefore 
be made available to the computer at one time for automatic 
use. The tapes may be fed in either direction past the magnetic 
heads. Any tape may be used for input or output at any time, 
and both input and output functions may be overlapped with 
computing. Automatic interlocks are used to ensure that only 
one tape functions as input at a given time and only one tape as 
output at a given time. 



204 COMPUTING SYSTEMS 

Instructions for rewind are of two kinds. One allows reuse of 
the same tape in the problem when called for by further instruc- 
tions, and the other sets an interlock which is not released until 
an operator changes the tape. The latter feature is of importance 
in handling large statistical problems where the data exceed 10 7 
characters, and also in avoiding reel-changing setup time between 
problems. Results recorded on magnetic tape can be printed 
on a special typewriter (with continuous paper feed) which is 
automatically controlled by the coding on the tape. 

Storage. High-speed and low-speed storage units make use 
of mercury delay lines and magnetic tape, respectively. The 
acoustic storage unit has 100 channels and an inherent capacity 
(counting the spaces between digit groups) of 91,000 binary 
digits. 

Of the 10 magnetic tape reels for which servo controls are 
normally obtainable, the user may allocate as many as he wishes 
to instructions, input, output, or temporary storage. These 
allocations may vary within the same problem if desired. Actu- 
ally, all the instructions needed for a great many different prob- 
lems could be put on one tape reel. 

Approximately 2,000 blocks, of 720 characters each, can be 
stored on one reel. Each input or output instruction transfers 
one such block of information at a rate of about 800 12-character 
groups per second. (This corresponds with a transfer rate of 
about 70,000 binary digits per second.) 

Within the high-speed storage, special transfer instructions are 
provided for handling two consecutive groups which are stored in 
a single acoustic channel. 

Arithmetic Operations. Addition or subtraction of a number 
is executed in less than 600 microseconds ; multiplication requires 
about 2,500 microseconds, and division less than 4,000 micro- 
seconds. (These times include allowance for obtaining the 
instruction itself from storage.) Either a 22-digit product may 
be obtained or an 11-digit product with proper round-off, accord- 
ing to the instruction used. 

Supervisory Controls. A typewriter keyboard and automatic 
printer are directly connected to the UNIVAC computer for 
supervisory purposes. This permits manual control of computer 
operations, modification of programs for test purposes, auto- 
matic printing out of brief indications of progress in long com- 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 205 

putations, etc. A panel of selector switches and indicator 
lamps is included; when a stoppage due to a failure to satisfy the 
automatic checking circuits occurs, the nature of the failure is 
indicated, and further diagnosis and remedy are facilitated by 
use of the selector switches. 

Pulse Code. The code for every character contains 7 binary 
digits. One is the check digit. Two others are used for indicating 
nonnumerical characters and are always zero for decimal digits. 
The last four correspond, in the case of a decimal digit, to the 
binary equivalent of the decimal number plus three. Thus, the 
last four digits of the code for 7 are 1010. In Sec. 13-4-5, this 
excess-three code is used as an example of one system for repre- 
senting decimal digits in binary notation and some properties 
of the system are mentioned. 

The check digit is always chosen to make the sum of the binary 
digits odd. Hence, the complete 7-digit code for 7 is 1001010. 
When the second or third digit from the left is nonzero, the 
code represents a nonnumerical character. The check digit is 
still chosen as before, to make the sum of the binary digits in 
the 7-pulse group odd. This checking code is used throughout 
the computer, both in high-speed circuits and on the magnetic 
tape. 

Checking. During all transfers, within the computer and to 
and from magnetic tape, special circuits examine every character 
(7-pulse group) to see that it is odd. If it is not, computer 
operation is arrested and an error indication given. Circuits 
for arithmetic operations are duplicated, and the outputs of the 
duplicate units compared. If the outputs are not identical, 
computation is stopped and an appropriate indication given. 
Periodically, every character group in the acoustic storage is 
scanned to check that its digits are odd, so that defects in the 
storage circuits will be found almost immediately even though 
the computing routine has not yet had occasion to call upon the 
defective register. The time required to carry out this periodic 
check is negligible. The time interval between such periodic 
checks is short enough to make the probability of making two 
compensating errors effectively zero. Some other checking 
methods are used in the tape recording device (UNITYPER) and 
the automatic printers (UNIPRINTERS). 

Alarm-type fuses (Western Electric), similar to those used in 



20G COMPUTING SYSTEMS 

the ENIAC, are used in all distribution lines for d-c power, and 
a blown fuse will cut off all d-c power. Problems can be so 
programmed that a shutdown, either from a blown fuse or from 
external power failure, does not require starting the problem 
over from the beginning. 

10-4-4. The SDC Raytheon Computer (The Raytheon Manu- 
facturing Company). The Raytheon Manufacturing Company 
has started to work on a high-speed digital computer 4 for the 
Special Devices Center of ONR. This machine uses mercury 
delay lines for internal storage and magnetic tape for external 
storage. 

Input and Output. Problems are introduced to the machine on 
coded magnetic tape. The code is generated by means of a 
manually operated keyboard. The magnetic tape can store 
either decimal numbers in binary code or binary numbers in the 
octal notation. Through the use of the octal notation, three 
binary digits identify each digital position. With the use of 
this number system the conversions from the binary to the octal 
and from the octal to the binary are made simple. Associated 
with the keyboard device is an input page printer, which prints, 
for visual observation and record, the information which has 
been introduced to the machine. 

The numerical output of the machine is recorded on magnetic 
tape and is then distributed to one of two output printers. 
The output data are normally in decimal form, although binary 
numbers can be printed in the octal notation. 

Storage System and Number Representation. The SDC Com- 
puter is capable of handling several different kinds of numbers. 
In standard operation it can store numbers to a precision of 
35 binary digits with fixed binary point. The basic arithmetic 
operations are carried out to a precision of 35 binary digits. It 
is also possible, however, to store and manipulate numbers of 
70 binary digits as pairs of standard numbers. Conversions 
from decimal to binary notation and from binary to decimal nota- 
tion are performed in the arithmetic unit and not by separate 
equipment. This requires that decimal numbers be stored in 
the internal, or acoustic-delay-line, system. The storage 
capacity per number in the storage system, which accepts num- 
bers in binary-coded decimal notation, is eight decimal digits. 

The internal storage will have a capacity of 1,152 numbers; 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 207 

each is represented by 35 binary digits with algebraic sign and is 
characterized by a check number. Three different types of 
operation are performed by this storage system: the writing of 
new numbers into the system, the reading of current information 
in the storage unit, and the erasing of information to prepare 
the internal storage for introduction of new numbers. With 
acoustic storage, numbers are circulated through the delay line 
in serial order and are regenerated by an electronic circuit which 
reintroduces them to the acoustic delay line. The pulse repeti- 
tion rate is 3.7 megacycles per second, and each number requires 
42 pulse positions. The time duration of one number is there- 
fore 11.25 microseconds. The acoustic delay lines, however, 
will store 32 numbers per channel so that a complete cycle 
through the acoustic delay line requires 360 microseconds. 
Thirty-six acoustic-delay-line channels are required to provide 
capacity for 1,152 numbers (32 with the main storage, plus a 
buffer reservoir housed with each of the four external storage 
units). The external storage consists of four identical self- 
controlled units each of which contains one magnetic-tape mecha- 
nism capable of storing 10 5 numbers. The tape on any individual 
external storage unit may be scanned at the rate of about 500 
numbers per second. From the design characteristics it may 
require, in the most unfavorable cases, as long as 2000 seconds 
to read a given number out of the external storage unit. This 
premium on time for reading from the external storage may 
require the rearrangement of data before they are placed into 
the machine so that the external storage unit presents recorded 
information at the right time and in the correct sequence for 
machine use. Each tape unit in the external storage unit has 
associated with it two 32-number acoustic delay lines, which 
are employed as buffers between the tape and the main electronic 
part of the machine. 

The transfer and storage of numbers in both the acoustic delay 
lines and the magnetic-tape storage unit are checked by means 
of a weighted count. This is a weighted sum of the digits of the 
number. This weighted sum is computed modulo 16 and is then 
modified by the addition of 1. A number to be checked and its 
weighted count cannot therefore be zero simultaneously, so 
that a null number is not a valid one. By this check, the failure 
of a gate or other device controlling the transfer will be detected. 



208 COMPUTING SYSTEMS 

Arithmetic Unit. The arithmetic unit of the SDC machine 
performs the arithmetic processes of addition, subtraction, mul- 
tiplication, and division. This unit can also generate new num- 
bers or new orders; the ability to generate new numbers differ- 
entiates the arithmetic unit from all the other units of the 
machine whose functions are purely passive. The arithmetic 
operations are performed by electronic-vacuum-tube techniques. 
The numbers in the arithmetic unit are handled by using a 
parallel representation, the various digits of the number being 
stored in separate flip-flop devices. This requires that the trans- 
formation of a number from storage to the arithmetic system 
be performed by series-to-parallel representation of the digits in 
each number. 

The arithmetic unit can also perform basic operations on either 
numbers or commands introduced to it. The latter initiate the 
transfer of numbers from one storage position to another, the 
shift of a number, and the extraction of any given number of 
digits from a number. 

The process of addition is performed in a straightforward way 
in this parallel representation. Subtraction uses the same 
process as addition after the complement of the subtrahend is 
taken. A convenient method for taking the complement makes 
use of the ability in the binary notation to replace all the digits 
in the subtrahend consisting of 1 by 0, and by 1. This com- 
puter takes complements on 2 — 2 -35 . Multiplication of two 
35-digit numbers results in the accumulation of the complete 
70-digit product in two of the arithmetic registers. Either 
35-digit section, or both sections, of this number can be used for 
subsequent calculation. In division, both the quotient and the 
remainder are available in two registers. 

All arithmetic operations are checked by means of an arith- 
metic weighted count so that the failure of any unit of the 
arithmetic system can be verified. 

10-4-5. ERA Computer (Engineering Research Associates, 
Inc.). Engineering Research Associates, Inc., also has com- 
pleted and submitted to the Bureau of Standards a preliminary 
design for a general-purpose electronic binary computer. 10 This 
design incorporates a magnetic drum 9 for the internal storage 
of numbers and orders. The general characteristics of the 
proposed machine include: punched-tape input, arithmetic, and 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 209 

coding operations 11 to be conducted by electronic vacuum tubes, 
binary digit representation, magnetic-drum storage, and punched- 
paper-tape or printed output. The machine is capable of 
handling 30-digit binary numbers and has a storage capacity of 
4,096 30-digit binary numbers. 

Input and Output. Numerical data and instructions for 
machine operation are introduced to the machine by coding 
on a seven-hole punched paper tape. This tape is prepared by 
manual operation of a keyboard translating unit which operates 
a standard seven-hole tape punch. The numerical data occupy 
six tracks on the tape; the coding for control of input operations 
occupies the seventh. For introducing a 30-digit number to the 
magnetic-drum storage, five consecutive lines in each of the six 
levels or tracks on the punched tape are used. Each track, 
therefore, supplies successively five digits to the drum. The 
30 digits on the magnetic drum may therefore be considered to 
be divided into six groups of five digits each. The first line or 
consecutive group of punches in each of the six tracks introduces 
a digit into the extreme left position of each of the groups of 
numbers into which the magnetic drum is divided. The second 
line or group of punches on each of the six tracks supplies digits 
to the digit positions immediately to the right of those previously 
filled. This procedure is continued until 30 binary digits are 
registered in a row across the magnetic drum. 

The punched tape is read by a photoelectric scanning mecha- 
nism capable of reading 150 lines per second. To introduce 
4,096 30-digit binary numbers requires approximately 2.3 
minutes. 

The seventh track, containing the control code for directions 
and for checking operations sequentially, is divided into con- 
secutive lines each composing a frame. These five lines are 
designated a, b, c, d, and e. Four phototubes read the seventh- 
track control code. These four tubes follow a binary coding 
which then identifies the operation directed by that line of tape 
perforations. 

The output section of this computer contains a unit for trans- 
mitting data from the magnetic storage to punched paper tape 
or to an electric typewriter for printing. The electric typewriter 
then prints characters or numbers in base 10; the perforator 
punches six digits as a line of data on six- or seven-hole tape. 



210 COMPUTING SYSTEMS 

Storage. This computer has storage for 4,096 30-digit binary 
numbers. These digits are stored on magnetic tape which is 
contained on the circumference of a cylindrical drum; the drum 
is capable of continuous rotation, and the magnetic representa- 
tions for and 1 are recorded, read, and erased while the drum is 
rotating. The drum contains two groups of 30 parallel tracks; 
each track can store 2,048 digits around the circumference of the 
drum. Each group is therefore equivalent to an array of 
numbers of 30 columns and 2,048 rows; this array of numbers 
may be thought of as being wrapped around the circumference of 
the drum so that the first row is in a position equivalent to row 
2,049. Each of the 2,048 rows in this array contains a 30-binary- 
digit number, and all the digits in a 30-digit number lie in different 
columns. The location of a number in this storage system, 
therefore, is performed by locating the row or box containing the 
number. A 12-digit binary number called the address char- 
acterizes this location. The leftmost binary digit identifies one 
of the two groups of tracks; the remaining 11 digits determine 
one of the 2,048 (2 11 ) rows or angular positions of the drum. 

There are, additionally, 11 tracks used for number location and 
two tracks for the timing and control of storage operations. 
The group of 11 tracks for location contains 2,048 permanently 
recorded 11-digit binary numbers, each number designating an 
address of one of the 2,048 rows. 

Above each of the 60 tracks is a magnetic head capable of record- 
ing, reading, and erasing the information contained on the track. 
As the drum rotates, the two groups of 30 heads read two 30-digit 
numbers simultaneously, the sequence of such patterns repeating 
themselves every drum revolution. Associated with the reading 
of the information of each track is a group of 60 storage-reading 
amplifiers for reading signals from the selected group of storage 
tracks and for causing the number read from a particular position 
to be transmitted to a specified destination in the computer. 
Storage-reading amplifiers for three sets of output gates direct 
the numbers to the X register, to the program-control registers, 
or to the print or punch register. 

Arithmetic Section. The arithmetic section performs the 
operations of addition, subtraction, multiplication, division, and 
transfer. This section contains three principal components: 
the X register (X), the Q register (Q), and the accumulator (A). 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 21 1 

The computer performs all its operations on binary numbers in 
the X and Q registers, capable of handling 30-digit numbers. 
Negative numbers are represented as complements on 2 30 — 1. 
In the accumulator, however, which has capacity for 60 digits, 
negative numbers are represented as complements on 2 60 — 1. 
The extreme left digit in each register and in the accumulator is 
used to designate the sign of the number in the register; a 
represents a positive number; a negative number is represented 
by 1. 

In the accumulator, the sum, difference, product, or remainder 
is formed, together with other arithmetic operations. The 
operation of the accumulator is essentially subtractive. This 
means that everjr number transmitted to the accumulator is 
automatically subtracted (modulo 2 60 — 1) from the number 
previously there. A number is added into A by transferring 
its complement to A. 

A number in A may be shifted to the left by as many places 
as required from 1 to 59. This can be done by a single command. 
If a number is shifted to the left five places, for example, the digits 
dropped from the extreme left end appear in the same relative 
order in the digit positions at the extreme right which have been 
left vacant by this shift. In this sense, the shift of position of a 
number in the accumulator is circular. 

Since the accumulator contains a 60-digit number and since the 
operation of the accumulator is essentially subtractive with 
respect to the number introduced, the accumulator contains 
input gates which convert the 30-digit numbers transferred from 
the X or Q registers to the appropriate kind of 60-digit number 
for induction to the accumulator. 

The Q register performs several arithmetic and logical opera- 
tions. In division, for example, the quotient is formed in Q, 
and in multiplication the multiplier is held in Q. Since it has 
bilateral communication with the accumulator, the Q register 
may also be used as a rapidly accessible storage for one 30-digit 
number. 

The X register is principally used to receive numbers trans- 
mitted from the magnetic-drum storage system to the arithmetic 
section. The X register has no adding or digit-positioning shift- 
ing property. The addend, the subtrahend, the multiplicand, 
or the divisor is held in X during the corresponding arithmetic 



212 COMPUTING SYSTEMS 

operations. The number in the X register or its complement 
may be subtracted into the accumulator. 

The arithmetic section also contains a unit known as the arith- 
metic shift counter for counting the number of places by which 
numbers are shifted in the Q register or in the accumulator. 
The arithmetic shift counter counts modulo 60. 

Program-control Section. The program-control section directs 
the execution of instructions required in the performance of a 
mathematical operation. This program of instructions is 
introduced to the machine and is contained in the magnetic 
storage system. The program instructions for each operation 
are expressed as an aggregate of 30 binary digits. In this form, 
of course, the digits may be stored and transmitted as numerical 
quantities. Sequential operations in this coding system may 
be seen by examining the structure of the 30-digit command 
number. For example, the right-hand 12 digits of a 30-digit 
instruction represent y, the execution address. This identifies 
the box in the magnetic storage system which contains the orders 
for carrying out a particular instruction. The next 12 digits 
to the left represent the program address which identifies the 
box containing the next instruction in the program. The 
remaining six digits to the left represent the command code. 
This code directs the machine to execute one of the 39 commands 
which the machine understands. 

Associated with the program-control section are three com- 
ponents which receive the three parts of the 30-digit command 
instruction. The execution address is received in the execution- 
address register. The program address instructing the machine 
on the next operation is received in the program-address register, 
and the command code of six digits in the command-translator 
switch. This switch energizes one of its 39 output leads appro- 
priate to a given six-digit code. 

The program-control section also contains an end-point counter 
whose primary function is the counting of the number of repetitive 
routine operations which have been undertaken in a program. 

The machine is capable of undertaking 39 operational com- 
mands. Without going into detail with respect to each of the 
39 commands, it is considered sufficient here to enumerate the 
number of commands in each of the principal categories. For 
example, there are 12 additive commands, seven commands 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 213 

relating to the transmission of data between the accumulator 
and the magnetic storage system, five commands relating to the 
transmission of data into and out of the Q register, six commands 
relating to the arithmetic processes of the arithmetic sequence 
control, five test commands, and four commands relating to the 
print/punch register and the termination of machine operation. 
The main program control and the arithmetic sequence control 
operate at a rate of 400,000 pulses per second. This basic fre- 
quency is supplied by a clock-pulse generator. The shifting of 
digit positions in the Q register and in the accumulator is per- 
formed at this rate. The addition cycle in the accumulator is 
3 cycles of the basic clock frequency, or 3 X 7.5 microseconds. 
A new number may therefore be transmitted to the accumulator 
every 7.5 microseconds, or a number which has been transmitted 
to the accumulator may be shifted in position after a time of 
7.5 microseconds. 

10-5. Applications of Large-scale Digital Machines 

Large-scale digital machines can be applied to the solution 
of a variety of mathematical problems relating to the fields of 
physics, chemistry, applied mathematics, mechanical and 
chemical engineering, statistics, astronomy, 41 and the biological 
sciences. The following is a broad classification of equations 
whose solutions may be obtained by digital techniques : 

10-5-1. Systems of Linear Algebraic Equations. These equa- 
tions occur frequently in problems relating to vibration, chemical 
analysis, the control of chemical processes in weather problems, 
and in statistical analysis. In this class are included tech- 
niques 26 ~~ 30,44 for finding characteristic solutions of matrix 
equations; 33 these matrix relationships occur in quantum- 
mechanical calculations and also in classical physics dealing with 
vibration of independent or coupled systems. 

10-5-2. Harmonic Analysis. The determination of coeffi- 
cients and terms in harmonic series has particular importance 
in celestial mechanics, in radiation problems, and in the deter- 
minations of crystal structure from x-ray-, electron-, and neutron- 
diffraction data. 

10-5-3. Mathematical Tables. Several examples of the use 
of large-scale digital machines in the preparation of mathe- 
matical tables may be found in References 27, 36, 40, and 41. 



214 



COMPUTING SYSTEMS 



00 
B 

03 

co 


d 

_o 

03 

u 

03 

a 
o 

d 


d 
o 
+2 

03 
M 

03 
ft 

o 
d 


d 

.2 

>-. 

03 

ft 

o 
d 


d 


u 

03 
ft 

o 
d 


d 


09 

Sh 

03 

ft 

o 
d 


a 


03 

!h 

03 

ft 

o 

CI 


d 
o 

e8 

tH 

03 

ft 

o 
d 


a 



03 

In 
03 

ft 

O 

d 


03 

d-d 

3 03 

s ^ 

O c3 
03 rt 

Sh hO 

ft 03 
W CD 

d -u 


In operation with- 
out final storage 
and terminal 
equipment 


-d 

a 

03 

.flg 

O o3 

I 1 

o 
o 


M 

d3 

o C 

d 

o >, 
o § 

c3 o <5 


d" > d* d o3 i 

03 O <U 03 !h P 

03 l! *h •-• O > 

^3^ d a£ 
tf a S-*. ft d^ 
pq 03 g * a -s * 

£ *=! .d i> -P -P> W 


d" 

0) 
03 

03 

< 


d 2> 

03 « 

-d t-q 

pq 03-d" 
s2 >» 

O ^ 


T3 

d 
d 
o 

Sh 

o 

60 . 

a ^ 
*?> 

o . 

it, s 

PH 03 
^ bJD 

IP 


M 

Ph 

o 

03 

pq 
i— i 


03 

d 

03 3 

So 

'< 6 

a 

ft § 

25 

(h 03 

£H 


e3 

c 

d 

H 

pq 


-d" 
d 
d 
o 

O 

60 . 
d o3 

•£> 

o . 

£ S 

-H ^ 

03 3 

^ Q 


OJ 

m 60 

a3 
i a 

J3 03 
S^ 

lb 

t3 
03 03 

nil 


S fl 
-3 o 

-O o3 

■c.a 

O be 


'co 

u 

<v 
> 

*3 
P 

T3 

> 
u 
03 
W 




















.3 

o 

!» 

o c 

o 03 

B 
"C 






03 

d 

03 
(3 

-d 

Sh 

o 

>> 

a 

Sh 


03 
C3 

d 

03 

d 

6 
>> 
a 


T3 

6 

d 

eq 

>> 

> 

03 




CO 




-d 

(H 

o 

d 
pq 

> 

03 


o 

> 

03 


>> 

U 
cs 

6 h 
•r o 

a-s 

Fh 03 
O £ 

»H C 

o> o 

2 ° 

'3 

pq 


pq 


PQ 


u 

H d 

<4H 'C 

O 03 

r-l ® 

o d 
~o « 

03 f-h 
U 03 
O O 

o X 


nq 
H 
pq 


>> 

!h 
03 
>; 

"3 

(H 

03 

(H 

03 

w 


pq 


a 

o 
O 
>. 

3 

o . 

d ft 

03 J; 

leH O 

2o 

f *-> 
% % 

M d 
° ft 


H 
pq 


' m 

u 

03 
> 

'3 
U 

■d 

(h 

> 

(h 
03 

K 




0) 



3 
o 
03 

a 

03 

a 

03 


0) 

o 

d 

03 

d 

03 9 

3 

'■5 ^ HH 

I°iH 

O "d 03 

3:3 2 

<! 2t3 

fl 03 

fcH O > 

^ O % 
PQ kh 


i 

03 

o 
d 

03 

d 
a* 

03 

02 ^ 

03 O 

3 Jj 

m3 

M » 

d 03 

go 
pq 




> 

O 
u 

03 

d 

ft 

o 
O 

hq 
H 
pq 

■>* 


-^ 

03 
03 

tH 

03 

w 


u 

03 

H 

03 
O 

d 

03 

d p 
o- o 

CO ^ 
03 d 

"-+3 o3 
§^ 

co a 

pq 

CD 


O 
g 

pq 


> 

•d 

o 

u 

03 

d 
ft 

a 

o 
O 
tl 
H 

pq 

00 


03 

-d 

Ih 

03 
> 
o3 

M 

C73 


-d 
d 

iq 
3 

C3 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 



215 



co 

3 

CO 


43 
J* 
< 

03 

« a 

p-* 


d 

.2 

'-£ 

3 

CO 

d 

o 
o 

a> 

d 

P 


d 
.2 
o 

CO 

d 


o 

d 

P 


d 

_o 

o 

d 

CO 

d 
o 

43 

d 
P 


d 
o 

o 

h 

CO 

d 
o 

0/ 

-o 
d 
P 


d 



d 

CO 

d 


h 

U 

-a 
d 
P 


d 



d 
■+-> 

d 



43 

T) 

d 
P 


d 


'-5 



d 

co 

d 



h 

43 

a 
P 


d 
_o 

CJ 
CO 

d 



43 

d 

p 


d 



d 

CO 

d 



h 

43 

d 

P 


t3 
43 

-a 


43 

?! 

03 

o 
h4 


43 
03 

T3 
V 

,fl 

4 
PQ 


T3 
0) 

03 

d 

'53 

<u 

T3 
0) 
X! 

o 
Eh 


43 

d ^ 

hi 

Pq 

5 - 

'■g-g 


T3 

d" 

43 
93 

< 

4 

pq 


u 

p 

CO 

"o 
_d 

"o 
>>-; 

"co hH 

43 o3~ 

> d 
P* 


.* 

'd 

h 

O 

1$ 
O^ 

^l 
^^ 
03 ^ 

•d « 
d rn 

p w 


O 

CO 
43 
"43 

M 

d 
<J 

CO 



hH 

co" 
pq 


d 


M 

_d 
'3 

CO 

o3 

CO* 

pq 





"3 g 

1* 


43 1 

•S O 

^^ CO 
43 T3 

"03 c 

fl to 

M M 43 

'CO d *g 

03 So 
« ^ 03 

^ d M 
og. S 


J 8 

11 

£'3 

c * 

O M 

o 3 




<! 
CO 
P 

43 

PQ 


pq 

CO 
P 

CO 

p 
d 

H 






S3 

O 

"2 o3 

43 •« 

> d 
"d 
ta- 


pq 
<j 

CO 
P 


Pq 

CO 
P 




ft 

03 
d S 

pq^ 

co pq 2 
%<\ 
«S 03 
£Pa> 


M 

c 
'-5 

o 

«| 

o d 

C3 43 

e3 

£ 

Pq 


0) 

o 

a 

03 

d 

h 

o 

>> 

a 

h 
<1 


is 

01 
CO 03 

/"») CO 

o 

> <u 

o3 n 


43 
u 
d 

03 

d 

(h 

o 
>> 

a 

h 
< 


43 

d 
o3 

d 

6 

>> 

g 




rt 

O 
>> 

03 


CO 

pq 


CO 

pq 




CO 

pq 
5? 


>> 

03 

S u 
•C 5 
ft-g 

t, 03 
o £ 

~ a 
43 o 

"3 
ffl 


1? 

o 43 

o d 

J "9 

eg! 

8 -3 

o o 
o "C 


o 
O 

hb 

s 

d 
o 

CD 

"£> 

03 


>> 

■4 J 

Si? 

•si 

d^ § 

•5 d.S 

s >£ 


.2 
"S 
_d 

"o 
>. 

"co 

CD 
> 

'd 
P 


2 

'0 

.a 

*o 

>> 

M 

0) 
> 

'd 
P 


"ol 
O 

O 
_^ 

43 
> 

"d 

p 


opq 
S5 s 

«<- >> d 

43^^ 

!-§ s 

d •= 03 


CO 

pq 


O 


fe a 


9) 

d 
IS 

03 

a 

"8 

a 

03 


O 

> 
Q 


d 
a 

E 
o 
O 

d 
o 
43 
,4 

O 

Q 
co 


u 

43 
ft 

a 



O 

"3 
'Si 

'q 

CO 

< 


O 

< 
> 
Q 

O 


a 



_co 

"8 

a 



>> 

'co u 

fe * 

"d §• 
P 


43 
ft 

a 



"oi 

'So 

5 

c3 

"c 

O 

O 

CO 


-d 
ft 


ft 

a 



a 

43 

d 

CO 

pq 


03 

d 
a 

a 


w 


C73 



< 
> 

S 

p 

6 

CM 



£S 



co O 
"1 ft 



13 03 

§s 

— CO 

a ^ 

^d 43 

.2 .a 

dH 






43 CO 



<5H 



216 COMPUTING SYSTEMS 

10-5-4. Propagation Problems. This field includes the solu- 
tion of total and partial linear differential wave equations where 
boundary conditions and damping may be important. In the 
propagation of electromagnetic waves, the solution of the wave 
equation is important for the design of antennae and waveguides 
of various geometries and shapes. For example, the solution of 
the wave equation is required to be obtained in elliptic cylinder 
coordinates. In this coordinate system the wave equation may 
be transformed into Mathieu's equation. The mathematical- 
tables project of the Bureau of Standards has recently calculated 
the characteristic solutions of Mathieu's equation in regions of 
useful application to physical problems. 

10-5-5. Flow Problems. This field of problems 32 - 36 - 38 ' 39 
includes the analysis related to the flow of compressible fluids in 
subsonic, transonic, and supersonic speed ranges. The solution 
of differential equations describing fluid flow is made increasingly 
difficult when consideration is given to turbulence and viscous 
effects in a compressible fluid. Because of the increasing 
importance of both the transonic and supersonic speed ranges, 
it is necessary that solutions of these equations be obtained in 
order to provide an interpretation of experimental phenomena 
and a guide for future research and development on high-speed 
flight and on the design and testing of missiles required to 
operate near or above the speed of sound. 

Closely related to the above field of problems is the solution 
of equations describing the flow of heat, either as a steady-state or 
a transient phenomenon, 44 in bodies of complicated shapes. 
Cooling problems connecting the heat flow through a surface 
with compressible fluid flow along it are closely related to the 
solution of problems of the utmost importance in the engineering 
and design of high-velocity missiles. 

10-5-6. Problems in Nonlinear Mechanics. Problems occur 
in the field of electrical and mechanical vibrations, in the study 
of shock-wave propagation, in stability studies, and in the 
behavior of control systems such as servomechanisms, torque 
amplifiers, etc., to which large-scale computers may be applied. 

10-5-7. Integral and Integro differential Equations. Solu- 
tions of differential equations describing boundary-value prob- 
lems may be performed more readily at times by transforming 
the differential equation into an integral equation. In certain 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 217 

cases the integral representation of the differential equation may 
assume a very simple form. The integral methods using Fourier 
and Laplace transforms will adapt certain types of differential 
equations to immediate solution. In dealing with more com- 
plicated physical problems, however, a machine calculation in 
numerical integration may be required. It may also be conven- 
ient to express a physical system as an integrodifferential equa- 
tion. Both types may require a sufficient number of points 
to be calculated as to necessitate the use of a large-scale digital 
machine. 

10-5-8. Evaluation of Multiple Integrals. It is frequently 
required in problems relating to physics, chemistry, astronomy, 
and other branches of science that the numerical value for double 
and triple integrals is desired. Methods for evaluating these 
integrals, essentially extensions of rules developed for single 
integrals, 34,35 may be used for machine calculation. In problems 
connected with statistical mechanics, it may be required to 
evaluate the incomplete gamma function between n — 15 to 30 
and infinity. For such an application a machine calculation 
might be made starting with certain initial values obtained in 
Pearson's tables. 

In the evaluation of a triple integral, it may be necessary to 
perform 1,000 additions of 10-digit numbers which have been 
individually computed to provide a value of the integral at 
points on a three-dimensional lattice. 

10-5-9. Statistical Problems. Many complex problems arise 
in the field of statistics. 31 These problems relate to personnel 
location, procurement of raw material, import and export 
data, census analysis, 42 actuarial problems, 29 and many others. 
Although this field of machine application involves the per- 
forming of the basic arithmetic operations on data submitted 
to the machine, it is also of prime importance that the resulting 
information be in a form which is capable of statistical analysis. 
The calculation required to transform the input data into final 
results is a straightforward problem on most large calculating 
machines since it involves presenting to the machine numerical 
data with orders to operate on those data with arithmetic proc- 
esses. Input devices from which data may be handled statis- 
tically include punched cards, punched tape, and magnetic wire or 
tape in the order of increasing speed of access of information. 



218 COMPUTING SYSTEMS 

10-5-10. Life-insurance Problems. This field of application 29 
for large-scale calculating machines has been chosen here as an 
example of the usefulness of such calculations applied to business 
problems. In the life-insurance business, large-scale machines 
would be useful on policy transactions where a large number 
of arithmetic and printing operations may be required dealing 
with interest, loans, dividends, cash surrender values, and 
extended insurance. In problems of this kind, the machine 
must be capable of retaining and consulting tabular data 
relating to insurance rates for various categories of insurance 
policies and for various age brackets. In cases where machine 
calculation is exceedingly rapid, the requirement for built-in 
tabular data is not so stringent. Uses for such machines include 
pension-plan calculations, the determination of premiums on 
and values of policies, and the determination of proposed new 
dividend scales. It is conceivable that the machine could cal- 
culate the required result quickly enough to obviate the necessity 
of building in great quantities of tabular data. In this case, the 
dividend rate, age of the insured, type of policy, etc., would be 
introduced to the machine as raw data, and the necessary cal- 
culations on each policy transaction would be performed by the 
machine. 

REFERENCES 
Design Characteristics of Digital Computing Systems 

1. Aiken, H. H., and G. M. Hopper, The Automatic Sequence Controlled 
Calculator — I, Electrical Engineering, Vol. 65, Nos. 8 and 9, pp. 384-391 
(August-September, 1946); The Automatic Sequence Controlled Calculator 
— II, ibid., No. 10, pp. 449-454 (October, 1946); The Automatic Sequence 
Controlled Calculator—Ill, ibid., No. 11, pp. 522-528 (November, 1946). 

2. Alt, F. L., A Bell Telephone Laboratories Computing Machine — I, Mathe- 
matical Tables and Other Aids to Computation, Vol. Ill, No. 21, pp. 1-13 
(January, 1948); A Bell Telephone Laboratories Computing Machine — 27, 
ibid., No. 22, pp. 69-84 (April, 1948). 

3. Bloch, R. M., Mark I Calculator in Proceedings of a Symposium on 
Large-Scale Digital Calculating Machinery (Jan. 7-10, 1947), Annals of 
the Computation Laboratory of Harvard University, Vol. XVI (Harvard 
University Press, Cambridge, Mass., 1948), pp. 23-30. 

4. Bloch, R. M., R. V. D. Campbell, and M. Ellis, The Logical Design of 
the Raytheon Computer, Mathematical Tables and Other Aids to Compu- 
tation, Vol. Ill, No. 24, pp. 286-295 (October, 1948). See also General 
Design Considerations for the Raytheon Computer, ibid., pp. 317-322. 

5. Burks, A. W., H. H. Goldstine, and J. von Neumann, Preliminary 
Discussion of the Logical Design of an Electronic Computing Instrument. 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 219 

Report prepared under Contract W-33-034-ORD-7481 between the 
Research and Development Service, Ordnance Department, U.S. Army, 
and the Institute for Advanced Study, Princeton, N.J. (June 28, 1946). 

6. Campbell, R. V. D., Mark II Calculator in Proceedings of a Symposium 
on Large-Scale Digital Calculating Machinery (Jan. 7-10, 1947), Annals 
of the Computation Laboratory of Harvard University, Vol. XVI (Harvard 
University Press, Cambridge, Mass., 1948), pp. 69-79. 

7. Eckert, W. J., The IBM Pluggable Sequence Relay Calculator, Mathe- 
matical Tables and Other Aids to Computation, Vol. Ill, No. 23, pp. 
149-161 (July, 1948). 

8. The Eckert-Mauchly Computer Corporation, The UN I VAC System 
(1947). Pamphlet prepared by The Eckert-Mauchly Computer 
Corporation, Philadelphia. 

9. Engineering Research Associates, Inc., Selective Alteration of Digital 
Data in a Magnetic Drum Computer Memory. Report prepared for ONR, 
under Contract N6onr-240 Task I with Engineering Research Associates, 
Inc., St. Paul, Minn. (Dec. 1, 1947). 

10. Engineering Research Associates, Inc., Summary of Characteristics, 
Magnetic Drum Storage Computer. Report prepared for the National 
Bureau of Standards, Department of Commerce, under Contract CST- 
10133 with Engineering Research Associates, Inc., St. Paul, Minn., 
Part I (Nov. 30, 1948). 

11. Engineering Research Associates, Inc., Examples of Coding, Magnetic 
Drum Binary Computer. Report prepared for the National Bureau of 
Standards, Department of Commerce, under Contract CST-10133 with 
Engineering Research Associates, Inc., St. Paul, Minn., Part II (Nov. 
30, 1948). 

12. Goldstine, H. H., and Adele Goldstine, The Electronic Numerical 
Integrator and Computer (ENIAC), Mathematical Tables and Other 
Aids to Computation, Vol. II, No. 15, pp. 97-110 (July, 1946). 

13. Goldstine, H. H., and J. von Neumann, Planning and Coding of Problems 
for an Electronic Computing Instrument. Report, The Mathematical 
and Logical Aspects of an Electronic Computing Instrument, Part II, 
Vol. I, prepared under Contract No. W-36-034-ORD-7481 between the 
Research and Development Service, U.S. Army Ordnance Department, 
and the Institute for Advanced Study, Princeton, N.J. (1947). 

14. Goldstine, H. H., and J. von Neumann, Planning and Coding of Problems 
for an Electronic Computing Instrument. Report, The Mathematical and 
Logical Aspects of an Electronic Computing Instrument, Part II, Vol. II, 
prepared under Contract No. W-36-034-ORD-7481 between the 
Research and Development Service, U.S. Army Ordnance Department, 
and the Institute for Advanced Study, Princeton, N.J. (1948). 

15. Hartree, D. R., Calculating Machines (Macmillan & Co., Ltd., London, 
1947). 

16. Hartree, D. R., The ENIAC, an Electronic Computing Machine, 
Nature, Vol. 158, No. 4015, pp. 500-506 (Oct. 12, 1946). 

17. Harvard University, Staff of the Computation Laboratory, A Manual of 
Operation for the Automatic Sequence Controlled Calculator, Annals of the 



220 COMPUTING SYSTEMS 

Computation Laboratory of Harvard University, Vol. I (Harvard Univer- 
sity Press, Cambridge, Mass., 1946). This manual contains an exten- 
sive bibliography on numerical analysis, pp. 338-396. 

18. International Business Machines Corporation, IBM Automatic Sequence 
Controlled Calculator (1945). Pamphlet of the International Business 
Machines Corporation, New York. 

19. International Business Machines Corporation, IBM Selective Sequence 
Electronic Calculator (1948). Brochure of IBM Corporation, New 
York. IBM Selective Sequence Electronic Calculator Is Dedicated to 
Science, Business Machines, Vol. 30, No. 11, pp. 1-12 (March, 1948). 

20. von Neumann, J., The Principles of Large-Scale Computing Machines 
(mimeographed by ORI, Navy Department No. 6553) 16 pp. Address 
delivered at a meeting of the Mathematical Computing Advisory Panel 
of the Office of Research and Inventions, May 15, 1946. 

21. Stibitz, G. R., Relay Computers, Report 171.1 R of the Applied Mathe- 
matics Panel of the National Defense Research Committee (February, 
1945). 

22. United States Air Force, Scientific Planning Techniques, Discussion 
Paper 1-DU, Project SCOOP (Planning Research Division, Director of 
Program Standards and Cost Control, Washington, D.C., Aug. 5, 1948). 

23. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, Report on the EN I AC, Technical Report I (University of 
Pennsylvania, Philadelphia, June 1, 1946). 

24. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, Functional Description of the EDVAC, Research Division Report 
50-9, a report of development work under contract W-36-034-ORD-7593 
with the Ordnance Department, Department of the Army (University 
of Pennsylvania, Philadelphia, Nov. 1, 1949). 2 vols. 

25. Williams, S. B., Bell Telephone Laboratories Relay Computing System 
in Proceedings of a Symposium on Large-Scale Digital Calculating 
Machinery (Jan. 7-10, 194-7), Annals of the Computation Laboratory of 
Harvard University, Vol. XVI (Harvard University Press, Cambridge, 
Mass., 1948), pp. 41-68. 



Applications of Digital Machines 

26. Alt, F. L., The Multiplication of Matrices, Mathematical Tables and 
Other Aids to Computation, Vol. II, No. 13, pp. 12-13 (January, 1946). 

27. Alt, F. L., A Bell Telephone Laboratories Computing Machine — II, 
Mathematical Tables and Other Aids to Computation, Vol. Ill, No. 22, 
pp. 81-84 (April, 1948). 

28. Andrews, E. G., and H. W. Bode, Use of the Relay Digital Computer, 
Electrical Engineering, Vol. 69, No. 2, pp. 158-161 (February, 1950). 

29. Berkeley, E. C, Electronic Machinery for Handling Information and 
Its Uses in Insurance, Transactions of the Actuarial Society of America, 
Vol. 48, Part 1, pp. 36-52 (1947). 

30. Bleick, W. E., Calculating Machine Solution of Quadratic and Cubic 
Equations by the Odd Number Method, Mathematical Tables and Other 



LARGE-SCALE DIGITAL COMPUTING SYSTEMS 221 

Aids to Computation, Vol. II, No. 20, pp. 321-324 (October, 1947). 
Method for possible application to Mark I Calculator or to the ENIAC. 

31. Bush, V., As We May Think, Atlantic Monthly, Vol. 176, No. 7, pp. 
101-108 (July, 1945). A high-speed (microfilm) selector of the Memex 
type, suggested in this article, has been constructed by Engineering 
Research Associates, Inc., for the Department of Agriculture. Pre- 
liminary tests indicate satisfactory performance. 

32. Clippinger, R. F., Airflow Problem Planned for the ENIAC, Mathe- 
matical Tables and Other Aids to Computation, Vol. Ill, No. 23, pp. 
206-207 (July, 1948). Summary of paper presented at the meeting 
of the Association for Computing Machinery held at Aberdeen Proving 
Ground, Aberdeen, Md., Dec. 11-12, 1947. 

33. Curry, H. B., and W. A. Wyatt, A Study of Inverse Interpolation on the 
ENIAC, Report 615 (Ballistic Research Laboratories, August, 1946). 

34. Harvard University, Evaluation of the Function 

S(byh) = Psin (x 2 + 6 2 ) 1 ' 2 dx/(x 2 + b 2 ) 1 ^ 

Report 2, Computation Laboratory Reports for the Bureau of Ships, 
Navy Department (June, 1944); Evaluation of the Function 

C(bih) = f h cos (x* + b 2 )^ dx/(x 2 + b 2 ) 1 ^ 

Report 10 Computation Laboratory Reports for the Bureau of Ships, 
Navy Department (October, 1944). 

35. Harvard University, Computation Laboratory, Fifteen-Place Table of 
e~ x , Publication 20, Computation Laboratory for the Bureau of Ships 
(July, 1945). Out of Print. 

36. Harvard University, Staff of the Computation Laboratory, Tables for 
the Design of Missiles, Annals of the Computation Laboratory of Harvard 
University, Vol. XVII (Harvard University Press, Cambridge, Mass., 
1948). 

37. Harvard University, Staff of the Computation Laboratory, Tables of the 
Modified Hankel Functions of Order One-Third and of Their Derivatives, 
Annals of the Computation Laboratory of Harvard University, Vol. II 
(Harvard University Press, Cambridge, Mass., 1945); Tables of the Bessel 
Functions of the First Kind of Orders Zero and One, ibid., Vol. Ill (1947) ; 
Tables of the Bessel Functions of the First Kind of Orders Two and Three, 
ibid., Vol. IV (1947); Tables of the Bessel Functions of the First Kind of 
Orders Four, Five, and Six, ibid., Vol. V (1947); Tables of the Bessel 
Functions of the First Kind of Orders Seven, Eight, and Nine, ibid., Vol. 
VI (1947); Tables of the Bessel Functions of the First Kind of Orders Ten, 
Eleven, and Twelve, ibid., Vol. VII (1947); Tables of the Bessel Functions 
of the First Kind of Orders Thirteen, Fourteen, and Fifteen, ibid., Vol. 
VIII (1947); Tables of the Bessel Functions of the First Kind of Orders 
Sixteen Through Twenty-Seven, ibid., Vol. IX (1948); Proceedings of a 
Symposium on Large-Scale Digital Calculating Machinery, ibid., Vol. 
XVI (1948). 



222 COMPUTING SYSTEMS 

38. Hicks, B. L., and H. G. Landau, Nonlinear Parabolic Equations, Mathe- 
matical Tables and Other Aids to Computation, Vol. Ill, No. 23, pp. 207- 
208 (July, 1948). Summary of paper presented at the meeting of the 
Association for Computing Machinery held at Aberdeen Proving 
Ground, Aberdeen, Md., Dec. 11-12, 1947. 

39. Holberton, J. V., Laminar Boundary Layer Flow in a Compressible 
Fluid, Mathematical Tables and Other Aids to Computation, Vol. Ill, 
No. 23, p. 208 (July, 1948). Summary of paper presented at the meet- 
ing of the Association for Computing Machinery held at Aberdeen 
Proving Ground, Aberdeen, Md., Dec. 11-12, 1947. 

40. International Business Machines Corporation, IBM Selective Sequence 
Electronic Calculator (1948) p. 3. Pamphlet of the IBM Corporation, 
New York. 

41. Irwin, J. B., The Expected Performance of the EDVAC on Some 
Astronomical Problems, Publications of the Astronomical Society of the 
Pacific, Vol. 60, No. 355, pp. 235-244 (August, 1944). 

42. McPherson, J. L., Applications of Large-Scale High-Speed Computing 
Machines to Statistical Work, Mathematical Tables and Other Aids to 
Computation, Vol. Ill, No. 22, pp. 121-126 (April, 1948). 

43. Mitchell, H. F., Jr., Inversion of a Matrix of Order 38, Mathematical 
Tables and Other Aids to Computation, Vol. Ill, No. 23, pp. 161-166 
(July, 1948). 

44. Seeger, R. J., On Computational Techniques for Certain Problems in 
Fluid Dynamics, in Proceedings of a Symposium on Large-Scale Digital 
Calculating Machinery (Jan. 7-10, 1947), Annals of the Computation 
Laboratory of Harvard University, Vol. XVI (Harvard University Press, 
Cambridge, Mass., 1948), pp. 157-167. 



CHAPTER 11 
ANALOG COMPUTING SYSTEMS 

11-1. Introduction 

Calculating machines of the analog type operate on physical 
quantities which can be measured, instead of on numbers which 
can be counted. When a problem expressed in terms of discrete 
digital quantities is to be solved by an analog machine, these 
quantities are translated from numerical into physical form, 
i.e., into light intensities, lengths, angular positions of shafts, 
electrical currents, and the like. Computing operations are 
performed on these physical quantities. The computed result 
also appears as a physical quantity which may be retranslated 
into digital form. 

Let us consider the concept of translation of digital data to 
physical form in terms of an example. Suppose the input to an 
analog computer is by manual settings of a continuously variable 
control dial. The operator turns the dial to the scale division 
corresponding to the number he wishes to insert. In setting the 
dial he rotates its shaft a number of degrees from the zero posi- 
tion. The angular position of the shaft is then a physical 
analog of the number it represents. It may, also, provide the 
means for setting a continuously variable rheostat and thus 
establish the value of a current. If so, the number on the dial 
is represented by the effective value of the variable resistance 
and by the amount of current flowing through it as well as by the 
position of the shaft. 

From this example it is easy to visualize an analog machine 
which receives its input information directly from some other 
machine, without an operator's intervention. Such devices 
are common. For example, the computer which solves the 
problem of a shell's trajectory and directs the aiming of an 
antiaircraft gun may take its inputs directly from the target- 
sighting apparatus. It should be noted that such a computer 

223 



224 COMPUTING SYSTEMS 

constantly solves the same problem, yielding output values which 
correspond with continuously changing input values. 

In discussing the question of relative freedom from error, 
as applied to computed results, it is well to start with a common 
understanding of the terminology. The word accuracy, as 
used herein, denotes conformity to fact. Precision denotes 
sharpness of definition. Thus, the statements ir = 3.1416 and 
7r = 3.141592G54 are equally accurate, but the latter is more 
precise. The first statement merely expresses the observation 
that the ratio of the circumference of a circle to its diameter is 
between 3.14155 and 3.14165 (as noted in Sec. 6-9). This 
statement is free from error; it conforms to fact. The latter, 
more precise statement (jr = 3.141592654) likewise conforms to 
fact. 

Suppose we are dealing with a computed result instead of with 
a direct measurement. Assuming complete freedom from arith- 
metic error, the precision with which we are justified in expressing 
the result depends upon the precision with which the input data 
were expressed and upon the nature of the functional relationship 
between the input and output data. 

In discussing the ability of a machine to yield computed results 
free from error, both of the terms we have just defined are useful. 
A computer is accurate which does not make mistakes. A digital 
computer yields highly precise computed results if it performs 
its operations accurately, if it carries a large number of digits 
in the register from which the answer is read, and, at the same 
time, is able to avoid round-off errors (see Sec. 6-12). Depending 
upon the nature of the functional relationships involved in a 
problem, the same computer may yield results which we are 
justified in expressing with greater or less precision. 

An analog computer is accurate if the relationships between its 
parts are valid analogs of the mathematical relationships in 
the problem. It yields precise computed results if its parts 
are made with precision and if the answers can be read with 
precision. The last part of this statement is particularly obvious, 
but nevertheless instructive. An analog computer yields an 
output in the form of a measurable physical quantity of some 
sort. If a measure of this quantity must be translated to digital 
form, the question of the precision with which the measurement 
can be read arises. The computer is similar in this respect to 



ANALOG COMPUTING SYSTEMS 225 

any other measuring apparatus. The result of the computation 
can be expressed no more precisely than it can be read. More- 
over, the entire operation of the analog computer is based on 
measurement. Though only the final quantity, the answer, may 
actually be read and translated into digital expression, other 
measurements are made by the interaction of cams, gears, cur- 
rents, etc., at each step in the solution, and the precision with 
which each of these measurements can be made depends upon the 
niceties with which the machine is fabricated and with which 
its components interact. 

In the case of the fire-control computer mentioned above, we 
have a machine from which we do not attempt to read results. 
We merely feed the results as controlling and directing orders 
to other equipment. The same problem is being solved con- 
tinuously; yet the input values are changing. Because this is 
true, the ability of the machine to follow changes becomes impor- 
tant. If there is an appreciable time lag between the incidence 
of a change in input. data and its effect on the output, the instan- 
taneous value of the computed result will be in error; i.e., the 
machine will be temporarily inaccurate. Generally speaking, 
such an analog machine's instantaneous accuracy depends upon 
its ability to follow these changes, a machine property which is 
unimportant when we are dealing with a static analog device like 
the slide rule. 

11-2. Components for Analog Computation 

11-2-1. Arithmetic Elements. Addition. In most mechanical 
analog machines, addition is performed by differential gears. 11 
There are several types of such differential gears, and since these 
have been described adequately elsewhere in the literature, it 
will suffice here merely to include the names of several of these 
types : the bevel-gear differential, the cylindrical-gear differential, 
the spur-gear differential, and the screw differential. In a 
differential gear adder the quantities to be added are represented 
by shaft rotations ; the sum of the augend and the addend is also 
a shaft rotation. For example, if it is desired to add X and Y 
and find the sum Z, X is the angular rotation of a given shaft 
from a predetermined zero position. For Y the rotation of a 
similar shaft on the other differential gear is used. The sum Z 
is the angular rotation of the shaft on a planetary gear which 



226 



COMPUTING SYSTEMS 



bears on both the gears to which the shafts of the quantities X 
and Y are related. 

Subtraction. Subtraction is performed generally by differen- 
tial gear mechanisms in the same manner as addition, except 
that the appropriate sign of rotation of the shafts is altered to 
provide the difference X — Y. 

Multiplication. There are several methods by which mul- 
tiplication may be performed with analog equipments. Mul- 
tiplication may be performed with a slide multiplier, with a 







ZERO POSITION 
OF 2 

Fig. 11-1. Slide multiplier. The bases (ki and X) and the altitudes 
(Z and Y) of the two right triangles are proportional because the triangles are 
similar. The arm forming the hypotenuse of the larger triangle is pivoted 
at their common vertex. 

nomographic multiplier, with a resolver, or by special adaptation 
of a wheel-and-disk integrator. Some analog machines contain 
both slide multipliers and integrators. 

The slide multiplier 11,32 is based on the principle of the propor- 
tionality of the sides of two similar triangles (see Fig. 11-1). The 
first triangle has a base of length h\ and an altitude Z; the second 
triangle has a base of length X and an altitude Y. By the rela- 
tionship between similar triangles, 

fci Z 



X 



or 



(XZ) 



(11-1) 



This type of mechanism has been used by the General Electric 
Company in the design of their differential analyzer. 22 

Multiplication can also be performed by an integrating 
mechanism; this is an adaptation of the Kelvin wheel-and-disk 



ANALOG COMPUTING SYSTEMS 



227 



integrator 9,32 (see Fig. 11-2). A sharp-edged wheel resting on a 
disk is driven by frictional contact with the disk; the axes of 
rotation of the wheel and of the disk are orthogonal. The radial 
position of the integrating wheel resting on the disk may be 
varied by a lead screw through the journal on which the wheel 
is mounted; this changes the relative speed of rotation of the 
shafts connected to the integrating wheel and to the disk. In 
this integrator there are two input shafts and one output shaft. 



JOURNAL 



SCREW FEED 

/ 



wT\nnnT\nnn^n^!\nn^^nnnnnnnnw;it\nnj 




INPUT 
Y SHAFT 



DISC 



INPUT 
X SHAFT 



Fig. 11-2. Wheel-and-disk integrator. Rotation of Z, in angular measure 
from an initial reference position, is proportional to fYdX. 

The input motion X is made to rotate the disk shaft, the angular 
rotation of this shaft representing the value of the quantity X 
at all times. Input motion Y rotates the lead screw controlling 
the radial position of the wheel on the disk. The output of the 
integrator is the shaft rotation Z of the wheel. If k 2 is a constant 
of the integrator, then, at any instant, the relationship repre- 
sented may be expressed as 

dZ 



dX 



= k»Y 



(11-2) 



or 



Z = k 2 fYdX (11-3) 

If two integrators are used, their combination can be made to 
provide the product XY. In the first integrator the variable X 
is represented by the rotation of the disk shaft, the variable Y by 
the rotation of the lead screw of the integrating wheel. In the 
second integrator the positions of X and Y are reversed. The 



228 COMPUTING SYSTEMS 

outputs Zi and Z 2 of the two integrators are then added to give 
the product XY. 

XY = Zi + Z 2 = fX dY + fY dX (11-4) 

The factor k 2 may be eliminated by adjusting the outputs of each 
integrator by this constant. 

Where the two variables appear within the integral, 

jXYd<f> 

two integrators will provide the required product. Since 

fXY d<j> = JX d((Y dcj>) = jX dS (11-5) 

the integrator, with 4> and Y as input variables, supplies the 
quantity 

S = JYdcf> (11-6) 

The variable S is then introduced to the second integrator as a 
shaft rotation of the disk, and X is introduced as the lead-screw 
rotation to the wheel carnage, so that the result 

R = fXYd4> (11-7) 

is generated by the output of the second integrator. 

For further details on special operations, such as the use of 
the wheel and disk as an inverse integrator, refer to Bush and 
Caldwell (pp. 316-318). 9 

Division. Division is performed either by the slide multiplier 32 
or by the integrator. 9 If the slide multiplier is used, the inputs X 
and Y are interchanged to give 



- *■ © 



(11-8) 



Division may also be performed with two integrators by taking 
the reciprocal of one of the variables, Y, for example, and using 
this reciprocal as the input to the Y (lead-screw) shaft in the first 
integrator and to the X (disk) shaft in the second integrator: 

j= f Xd(^j + f^dX (11-9) 

11-2-2. Differentiation and Integration. Both differentiation 
and integration are performed with the Kelvin wheel-and-disk 



ANALOG COMPUTING SYSTEMS 229 

integrator. The method of its use is described in Sec. 11-2-1 
above under Multiplication. 

11-2-3. Input and Output. Numbers and orders to a differen- 
tial analyzer may be introduced by means of coded punched tape. 8 
The orders for machine setup include those relating to the shaft 
connections and to the gear ratios required for the solution of a 
given problem; numbers representing the initial conditions may 
also be introduced on punched tape, or in the case of earlier units, 
these may be set into the machine manually. 

Where input of a function in graphical form is required, an 
input table 9>22 ' s2 is used. This is usually a flat platform on which 
the required function is plotted in cartesian coordinates. The 
input table is equipped with a pointer for following the plotted 
function; if f(x) is plotted along the ordinate and x along the 
abscissa, the ordinate position is often controlled manually by a 
feed-screw mechanism, while the traverse along the abscissa is 
driven by a motor. The x drive on the input table is taken from 
the basic x drive of the machine and must always be in syn- 
chronism with it so that the correct value of fix) is introduced 
to the other elements (adders, multipliers, integrators, etc.) in 
the proper phase. Automatic curve-following devices have been 
reported by MIT 18 and the University of California at Los 
Angeles. 32 

The output of a differential analyzer may be punched on paper 
tape or printed in tabular form by an electrical typewriter. 
Frequently, it is convenient to plot the output graphically. An 
output table, 9 ' 22 '* 2 similar in design to the input table described 
above, is used for this type of output and consists of a flat table 
equipped with a frame for support and manipulation of a stylus. 
The function fix), or y, is plotted automatically on the ordinate 
and the independent variable, x, on the abscissa. 

11-2-4. Generation of Functions. Functions in graphical form 
may be introduced to the machine through the input table. 
However, when certain simple analytic functions are required as 
elements in the solution of a differential equation, they are fre- 
quently generated within the analyzer itself. 

The square of a function may be generated by one integrator. 
Since y — x 2 , y may also be expressed as 

y = f2xdx (11-10) 



230 COMPUTING SYSTEMS 

2x is introduced to the lead screw on the wheel carriage and x to 
the disk shaft. 

The function y = e x may be generated with one integrator by 
feeding the output (rotation of the integrating wheel) back into 
the shaft on the carriage, since 

y = e x ; j- = e x ; y = J e x dx and y = fy dx 

(11-11) 

By feeding y = 1/x into an integrator z = In x can be gen- 
erated. If n times the output, nz, is fed into another integrator, 
the function y = x n is generated, since 

i 7 dx 

ay = nx n ~ l dx = nx n — 

dy = yd(n In x) (11-12) 

y = fy d(n In x) 

In this operation, the output y of the integrating wheel is fed back 
as the input to the lead screw on the wheel carriage. 

For a more detailed discussion of these connections, see articles 
by Amble, 1 Bush, 8 Bush and Caldwell, 9 Hartree, 15 and Shannon. 30 

11-3. The MIT Differential Analyzers 

The first large-scale differential analyzer 8 was constructed at 
the Massachusetts Institute of Technology and put into operation 
there in 1930. The basic design of this machine followed explora- 
tory work at MIT on numerical methods 5,6 during the period 
1925 to 1930. The original differential analyzer operated on 
mechanical principles. This machine has been duplicated at a 
number of laboratories in this country and abroad. 

In 1945 a new type of differential analyzer, completed and 
put into operation during the war at MIT, was announced by 
Bush and Caldwell. 9 The new machine made automatic many 
of the manual operations of the earlier mechanical model and 
increased the capacity and precision. It is appropriate here to 
discuss the principal features of this machine in terms of the 
types and number of components it contains for computation. 

11-3-1. Input and Output. Numbers and orders are intro- 
duced to the machine 9 by coding on three punched-paper-tape 
units. These tapes are designated A, B, and C. The A tape 



ANALOG COMPUTING SYSTEMS 231 

supplies directions for interconnecting the computing units of 
the machine. The B tape is punched with gear-ratio data; 
the C tape contains the initial numerical data with which each 
of the computing units must be supplied to start the solution 
of a given equation. Each tape is punched with data identifying 
the problem under investigation. 

The A tape, with the interconnection data, uses a four-digit 
decimal code to describe the shaft acting as a source (transmitter) 
and the shaft which is driven (receiver). The first two digits of 
the code designate the type of shaft to be used; the second two 
digits specify which one of the computing elements possessing 
such a shaft is to be chosen. For example, in a hypothetical 
case, the digit group 2802-3410 might specify that the output 
shaft (code 28) of the second integrator (code 02) is to drive an 
input shaft (code 34) on the tenth gearbox (code 10). Each 
transmitter-receiver connection appears on a frame of the coded 
tape. 

The B tape, with its punched code representing the gear-ratio 
orders, contains numerical gear-ratio data and the numbers 
identifying the particular unit to which each order and number 
applies. 

The C tape carries coded numbers representing the initial 
conditions of the problem and the coded identification of the units 
which will make use of these numerical conditions. 

For input of graphical data, a modified form of function unit 
has been developed. The plotted function is mounted on a drum ; 
the drum is rotated by the input, or in accordance with the 
independent variable, while the pointer, or index, is kept on the 
trace of the function by manual rotation of the lead screw which 
moves the pointer along the direction of the axis of rotation of 
the drum. Since the output of the function unit must be a 
source of torque for relaying its information to other elements of 
the machine, the operator supplies this torque while acting to 
keep the index on the function. 

The output of the differential analyzer is a measurement of 
the rotation of shafts representing the desired variables. For 
example, if the dependent variable, y, is the result of integrating 
a differential equation with x as the independent variable, the 
output of the y shaft can be made to control the rotation of the 
screw feed geared to the stylus on the output table while the main 



232 COMPUTING SYSTEMS 

drive which rotates the drum is supplied with torque by the x, 
or independent-variable, shaft. 

For numerical registration of the shaft rotation, a counter 
unit is servoconnected to each shaft whose output is desired. 
Shaft rotations may be measured simultaneously on a number of 
units at any time and without stopping the machine. These 
data are stored in groups of relay counters. The system can 
handle up to ±10,000 revolutions and can record to 0.1 revolu- 
tion. The measurement of shaft rotation is first corrected for 
its initial value by a relay matrix capable of performing addition 
and subtraction ; the corrected data are then transferred from the 
matrix to storage relays prior to introduction to the automatic 
typewriters. The printing unit is automatic motor-driven IBM 
equipment. A set of 10 seven-digit numbers (with spaces and 
algebraic sign) can be printed every 12 seconds. 

11-3-2. Integrators. The new differential analyzer 9 has 18 
integrators at present; this number of integrators may be 
increased to 30. In view of the possible expansion to 30 integra- 
tors, the machine has already been arranged to operate as three 
sections, each having its own recording channel, typewriter 
units, computing matrices, decoding relays, and typewriter 
storage relays. 

The disk of the integrator is constructed of plate glass ground 
flat; the integrating wheel has a steel hub with a magnesium 
flange and a hardened-steel tire finished to a sharp edge (with 
straight sides, 60° included angle). The edge of the wheel bear- 
ing on the glass disk is ground to a width of 0.002 inch. 

Each integrator has an automatic setting mechanism which 
consists of a group of rotary selector switches for adjusting the 
integrator to the required initial position in conformance with 
coded orders on the C tape. The position of the integrating 
wheel may be set on the radius of the disk to a precision of 1 part 
in 150,000. The setting mechanism through its 44 associated 
relays may also be controlled centrally to read the position of 
the lead screw on the wheel carriage and transmit the data to a 
printing station. If the integrating-wheel displacement is to 
be multiplied or divided by a constant, e.g., 2, 5, or 10, the setting 
mechanism reads the wheel position and transmits the reading 
to a relay multiplier-divider unit, which performs the required 
arithmetic operation. The product or quotient is then trans- 



ANALOG COMPUTING SYSTEMS 233 

ferred back to the setting mechanism, which automatically 
adjusts the integrator to the correct position. 

11-3-3. Gear Assemblies. A problem on the differential 
analyzer is set up and ready to operate when each of the required 
shafts is connected to a source of torque. On the mechanical 
differential analyzer, the interconnection of these shafts was made 
manually. In the new machine, clusters of gears are assembled 
into gearboxes provided with electromagnetic shifting devices for 
change of gear ratio. The gearboxes are of two types: a one- 
digit decimal box provides any gear ratio between 0.1 and 1.0 in 
steps of 0.1; a decade decimal box provides any gear ratio 
between 0.0000 and 1.1110 in steps of 0.0001. The gearboxes 
are each driven by a standard servomotor; the output of the 
gearbox is registered by an angle indicator. 

11-3-4. Machine Control. The machine 9 is controlled through 
an automatic electrical switchboard which regulates the in- 
terconnection of data-transmission elements. Servomecha- 
nisms 3 ' 4,7 ' 10 ' 12,17,20 provide the power for driving all mechanical 
loads. Instead of measuring shaft rotations by selsyns, 24 
through magnetic-field variation, an angle-indicator unit operat- 
ing on the variation of an electric field was found to be feasible ; 
this unit operates at low torque and is capable of being used in 
conjunction with the standard telephone switching equipment 
for the transmission of data at low energy levels. 

The angle indicator 9 contains two capacitance bridges, each 
excited with voltage at about 3,000 cycles per second. Each 
bridge has a pair of fixed and a pair of variable condensers. The 
angular position is parallel. Rotating between the two plates 
of the variable condenser are a pair of specially cut cam-shaped 
disks of dielectric material; these are fastened to the shaft whose 
rotation is to be measured. As the shaft is rotated, each bridge 
undergoes a change in capacity such that the output voltages of 
bridge 1 and bridge 2 are 

ei = A;#isin <f> (11-13) 

e 2 = kE 2 cos cf> 

where e\ = output of first bridge 

e 2 = output of second bridge 
Ei and E 2 = exciting voltages 

<f> = angle of shaft rotation 

k = a constant of the system 



234 COMPUTING SYSTEMS 

The outputs d and e 2 may be used to transmit the position 
of one shaft A to another shaft B. In this case, the two bridges 
of A are excited with the same alternating voltage E; the output 
voltages ei and e 2 are transmitted through a switchboard and are 
introduced as input voltages to the angle indicator on shaft B. 
Since the output voltages from A are 

ei — kE sin 4> A 
and (11-14) 

e 2 — kE cos 4>a 

the output of the capacitance bridges associated with shaft B 
will be 

e z = kei cos 4> B ,-- 1 _ 

e 4 = ke 2 sin cf> B 

Combining Eqs. (11-14) and (11-15), 

e 3 = k 2 E sin (f> A cos <^ 5 , 

e 4 = k 2 E cos <^>a sin <f> B 

The output voltages e 3 and e 4 from angle indicator B are then 
subtracted in a transformer network. The transformer output 
is then 

e = e 3 - e 4 = /c 2 # sin (^ - B ) (H-17) 

The voltage e, for small differences in shaft positions, is propor- 
tional to the difference between the angular positions of the 
shafts A and B; this voltage is made the input to a servomotor 
which supplies the torque to control the shaft B. 

The servomotors used are of the repulsion type with two sets 
of brushes, one to provide positive, the other to provide negative 
rotation. The torque delivered by the servomotors is controlled 
by varying the plate current through each of a pair of thyratrons 
which are connected to the brush pairs through step-down 
transformers. 

The three different sections of the machine are each equipped 
with an independent variable drive; any one of these drives, 
however, may be used with all or any part of the computing 
elements of the entire machine. The power element for the 
independent variable drive is a variable-speed motor with an 
associated angle indicator. This angle indicator is associated 
with a primary element for machine control; it can drive any num- 



ANALOG COMPUTING SYSTEMS 



235 



ber of receiving angle indicators (11 have been so driven). The 
speed and acceleration of this shaft are controlled automatically. 
11-3-5. Relay Equipment. Several thousand relays are in 
use at present in the new differential analyzer. These relays 
are of the telephone type. They are used for decoding and 
executing input-tape orders, controlling the electrical typewriter 
output, performing addition and subtraction in the registration 
of shaft rotations, multiplication and division for integrator 

Table 11-1. Table of Equipment 
The new differential analyzer contains the following equipment: 

Independent variable units 3 

Integrators and setting mechanisms 18 

High-speed numerical counters 10 

Numerical recording channels (including special typewriters) 3 

Manual function units (these units can also be used for graphical record- 
ing) 3 

Graphical recording unit 1 

Automatic function units (not yet in operation) 4 

Gear combinations — all gearboxes can be used separately or with asso- 
ciated adder gears. The details of gearing available are given in the 
following table. 

Table 11-2. Gear Table 



Type of unit and 
gearboxes contained 


No. of 
units 


Total no. of 
decade boxes* 


Total no. of 

one-digit 

boxes t 


Total no. 
of adders 


3-input adder unit, with 
2 decade boxes and 1 
direct drive 


2 
6 
3 
3 
3 


4 
6 
3 

3 



6 

3 
3 


4 


2-input adder unit, with 
1 decade box and 1 
one-digit box 


6 


2-input adder unit, with 
1 decade box and 1 
direct drive 


3 


2-input adder unit, with 
1 one-digit box and 1 
direct drive 


3 


Nonadding unit, with 1 
decade box and 1 one- 
digit box 









Grand totals 


17 


16 


12 


16 







* Decade boxes provide all ratios from to 1.1110, in steps of 0.0001. 
f One-digit boxes provide all ratios from 0.1 to 1.0, in steps of 0.1. 



236 COMPUTING SYSTEMS 

positioning, counting of shaft rotations, and for temporary 
storage. 

Tables 11-1 and 11-2 summarize the equipment in the new 
differential analyzer at MIT. These data were presented by 
Bush and Caldwell (p. 326). 9 

11-4. The General Electric Differential Analyzer 

An improved type of differential analyzer has also been 
designed and built by the General Electric Company. 22 One 
model of this analyzer is in current use at the University of 
California at Los Angeles. 32 

The general mechanical features of the components of this 
analyzer differ in several respects from those in the new differen- 
tial analyzer at MIT. For example, multiplication and division 
are performed generally with a mechanical multiplier by using 
similar-triangle relations (see Multiplication in Sec. 11-2-1), 
although the two-integrator method for obtaining products may 
still be involved. 

The integrator output shaft is not in physical contact with the 
integrating wheel but is controlled by a photoelectric follow-up 
system. Since it is not connected to the output shaft, the wheel 
experiences no reaction from the torque drive of that shaft; the 
only torque on the wheel itself is that imposed upon its own 
bearings. The integrator wheel is a Polaroid disk with a steel 
rim and hub; the wheel itself forms part of the optical system. 
A split beam of light passes normally through the wheel face, 
one beam traversing one of two additional disks and the other 
beam traversing the second disk. The two disks are set with 
their planes of polarization at 90° to one another; they are 
mounted on a common shaft geared to a follow-up motor. The 
shaft of this motor simulates the rotation of the integrator 
wheel. 

After passing through the integrator wheel and one of the 
polarized disks, each beam falls upon a photocell. As the 
integrator wheel rotates, the illumination on the two photocells 
is varied. The photocells form two arms of a d-c bridge; the 
output, or current difference between the photocells, causes one 
thyratron to conduct more and another thyratron to conduct less 
than its previous current output. The thyratrons both supply 
current to a pair of field coils on a split-series servomotor. The 



ANALOG COMPUTING SYSTEMS 237 

sign of the difference in the current outputs controls the direction 
of rotation of the motor. Therefore, an unbalance in the light 
striking each photocell is transmitted by thyratron control so as 
to correct this unbalance by changing the angular position of the 
polarized disks. This method of balancing is used to keep the 
shaft of the servomotor at the same angular position as the shaft 
on the integrator wheel. 

Table 11-3 gives a list of components in the differential analyzer 
at the University of California at Los Angeles; this material is 
presented in Reference 32 (p. 2). For the design characteristics 

Table 11-3. Components of the Differential Analyzer 9 at the 
University of California at Los Angeles 

Adders (also subtract) 20 

Multipliers (also divide) 3 

Integrators 14 

Input tables 4 

Independent variable motors 2 

Output tables 2 

of other differential analyzers, see articles by Beard, 2 Hartree, 13 ' 16 
Hartree and Porter, 14 Lennard- Jones, Wilkes, and Bratt, 23 
Massey, Wylie, and Buckingham, 25 and Murray. 26 

11-5. Electronic Differential Analyzer 

Recently attention has been given to performing differentia- 
tion, integration, and arithmetic operations with electronic 
circuits. 19,21 Analogous mechanical methods have been outlined 
previously, and these preceded the development of electronic 
units for differential analysis. The basis for all the electrical 
analog computers is the mathematical relation between currents 
and voltages in electric circuits. For example, the voltage drop 
across a resistor is proportional to the algebraic sum of the 
currents which are introduced from various sources. Thus, 
KirchhofFs law is the basis for an electrical adder. 

In general, the order of magnitude of error in an analog com- 
puting device is closely associated with the ratio of output 
energy produced to input energy required. For this reason, the 
electron tube is a valuable aid to electrical computing because it 
requires very little input energy. The use of voltage or current 
amplifiers 29 in an electrical computer is somewhat analogous to 
the use of torque amplifiers in mechanical computers. 



238 COMPUTING SYSTEMS 

In the electronic differential analyzer the arithmetic operations 
are performed in the following way: Addition and subtraction 
can be accomplished through the use of two identical triodes in 
parallel in a resistance-coupled amplifier: The plate or output 
potential can be made equal to the sum or difference of the input 
(or grid) potentials multiplied by the factor 

— iiRl 
Rl ~1~ Rp 

where n = amplification factor of tube 
Rp = plate resistance of tube 
Rl = load resistance 

Multiplication by a constant factor is performed by a single 
stage of a resistance-coupled amplifier. 28 With proper adjust- 
ment of the operating point, the plate potential is a linear func- 
tion of the input or grid potential multiplied by the factor 
— IjlRl/(Rl + Rp.) It should be noted that this factor can be 
greater or less than 1. Thus, the circuit can be used to perform 
division as well as multiplication by a constant factor. 

Multiplication by a variable factor may be performed by 
introducing a voltage which is proportional to the multiplier 
across the outside terminals of a potentiometer. The poten- 
tiometer shaft is adjusted to an angle proportional to the multi- 
plicand, and the output voltage is proportional to the product 
of the two input quantities. Since both input voltage and shaft 
angle may be variable quantities, the product of two variable 
quantities may be formed. 

Varying electrical quantities can be generated with an input 
table similar to the ones used in mechanical differential analyzers. 
In the case of an electrical machine, the shafts which drive the 
curve-following stylus are connected to potentiometers and cause 
an output voltage to vary in accordance with a plotted function. 
If a function is to be used frequently, a potentiometer can be 
wound on a card which varies in width in such a manner that the 
output voltage is the desired function of the shaft rotation. 
This device is the electrical equivalent of a mechanical cam. 

In order to perform differentiations and integrations, the 
properties of RC circuits may be employed. An RC circuit in 
conjunction with an electronic amplifier can be made to have a 
time constant of about 2 minutes. Thus, an integration could be 



ANALOG COMPUTING SYSTEMS 239 

performed for only approximately 3 seconds before a 5 per cent 
error caused by the leakage of charge from the capacitor is 
accumulated. By adding a feedback 27 circuit in the amplifier, 
the time constant of the RC circuit can be increased to about 
100 minutes, and integration may be performed for 5 minutes 
before a 5 per cent error is accumulated. 

A one-stage conventional resistance-capacity amplifier repre- 
sents a first-order differential equation with constant coefficients. 
The plate potential can be shown to be related to the grid poten- 
tial by a first-order differential equation with a time constant 
[R G + R l Rp/(Rl + Rp)]C, where C is the value of the coupling 
condenser and R G is the grid resistor in the next stage. 

Using n stages of resistance-capacity amplification, equations 
of the nth order can be simulated, and with the proper choice of 
circuit constants and tubes, differential equations with real 
negative coefficients can be set up. By using the principle of 
feedback the range of coefficients can be extended greatly. Also 
by using the feedback principle, a differential equation can be 
set up which has an auxiliary equation with complex roots 
whose real parts are either positive or negative. 

The output potentials of the machine are measured by means 
of a recording vacuum-tube voltmeter, and it has been suggested 
that a self-balancing potentiometer of the Brown and Leeds- 
Northrup type can be used to drive a plotting table. 

A commercial electronic differential analyzer is being produced 
by the Reeves Instrument Corporation, and is called the REAC. 33 
The basic machine contains 20 d-c amplifiers, of which seven are 
integrating amplifiers, seven are inverting amplifiers, and six are 
summing amplifiers. Servomechanisms are used to carry out 
multiplication of variable quantities, to resolve vectors, and to 
introduce arbitrary functions. Provision is made to couple 
several REAC's together in order to increase the capacity. A 
single machine will handle differential equations up to the 
seventh order. 

A circuit which can be used as a square or square-root com- 
puter element (as well as in various other applications) has been 
reported by R. D. Campbell. 68 It comprises a d-c amplifier, a 
feedback loop, and a diode, operated in the region where the slope 
of the plate-current-filament-current characteristic is relatively 
steep. The feedback is designed to maintain constant power in 



240 COMPUTING SYSTEMS 

the diode filament. When the circuit is to be used as a computer 
element, the input may be an a-c current fed through the filament. 
The output is the output of the feedback amplifier. The power 
generated in the filament is the sum of the mean squares of the 
input and output currents times the resistance of the filament. 
This is kept constant. Assuming the filament resistance is con- 
stant, which is a valid assumption so long as the frequency of the 
input current is not higher than about 10 megacycles, 

n + n = k 

where Ii and I are the input and output currents. There- 
fore K — I 2 varies as the mean square of the input current. 
This relation can be utilized in various computing applications 
with an accuracy of about 99 per cent. 

11-6. Electrical Network Analyzers 

Electrical network analyzers for solving algebraic equations 
and linear and nonlinear differential equations have been devel- 
oped by manufacturers of electrical equipment, principally the 
General Electric Company and the Westinghouse Electric 
Corporation. They were originally built for studying network 
problems in connection with the transmission of electrical power 
and voltage-current characteristics in systems where power loads 
and power sources may vary rapidly with time. In recent years, 
more general use has been found for these machines. The types 
of network analyzers include the d-c network analyzer, the 
a-c network analyzer, and the transient analyzer. These are 
described briefly below. 

11-6-1. D-C Network Analyzer. The D-C Network Analyzer 
has been in service for the solution of industrial problems for the 
past 25 years. This analyzer is a system of coupled resistance 
circuits which can be flexibly arranged to simulate the steady- 
state conditions existing in an electrical or mechanical problem. 
The voltage between various points of the resistance lattice 
and the current through any given group of resistance elements 
can be determined by direct measurement. Steady-state solu- 
tions for compressible fluid flow and heat flow in addition to 
solutions of Laplace's equations in two dimensions have been 
obtained by the D-C Network Analyzer. 36 In most problems of 
this type, the voltage across a resistance element simulates the 






ANALOG COMPUTING SYSTEMS 241 

electrical potential or temperature and the current flowing in the 
resistance element simulates the fluid flow or heat flow in the 
medium which is represented by this network. It is obvious 
that the precision of the results depends upon the precision with 
which electrical measurements can be made. Thus, a precision 
of 1 or 2 per cent is considered acceptable. 

11-6-2. A-C Network Analyzer. The design and development 
of the A-C Network Analyzer followed chronologically that of the 
D-C Network Analyzer. The first large-scale A-C Network 
Analyzer was constructed in 1929. This analyzer is similar to 
the d-c type with each resistance element being replaced by a 
combination of resistance-capacity and -inductance elements. 
It is more flexible than the D-C Network Analyzer since it can 
represent phase relationships between currents and voltages in 
various parts of the network and is therefore capable of handling 
complex quantities directly. The propagation of electro- 
magnetic waves has been studied with this analyzer, and during 
the Second World War it was used to investigate the char- 
acteristics of cavity resonators. In the field of mathematical 
physics the A-C Network Analyzer has been used to obtain 
solutions of the Schrodinger equation for a simple harmonic 
oscillator and for the rigid rotator in one dimension. Investiga- 
tions have been made on the effect of the shape of various 
potential wells on the eigenvalues and eigenfunctions for both 
the linear harmonic oscillator and the rigid rotator. 

11-6-3. Transient Analyzer. This analyzer differs from the 
d-c and a-c analyzers described above in that it can be used to 
study nonrepetitive phenomena. It was designed primarily 
to study electrical transient phenomena resulting from switching 
and voltage or current discontinuities. The effect of nonlinear 
elements in physical systems can also be investigated with this 
analyzer. 

The Transient Analyzer 36 is capable of having its circuit 
parameters varied easily. Therefore it is a powerful tool for use 
in the synthesis of an optimum physical system from a given set of 
components of undetermined value. For example, the char- 
acteristics of a servomotor and its load might be simulated, and 
the transient response of the over-all system observed as various 
servoamplifier circuit parameters are adjusted. The transient 
response at various points in the system may be observed on a 



242 



COMPUTING SYSTEMS 



cathode-ray oscilloscope, and a great deal of the design work can 
be done by bringing the response curves into the desired form. 
Thus, a minimum of output data needs to be permanently 
recorded for further analysis. The application of this analyzer 
to other nonlinear problems in electricity, mechanics, heat, and 
sound is expected to produce useful and informative results. 

11-7. Special Analog Systems 

11-7-1. Solutions of Simultaneous Linear Algebraic Equations. 

Analog devices on which linear simultaneous algebraic equations 




r 



Ovx' 



TAEE I 




a 



O Q 



TAPE 2 



Fig. 11-3. Wilbur's mechanism. The distance between the two ends of 
each tape is reduced by an amount proportional to a n xi + a\ 2 x 2 or a 2 \X\ + 
a 22 x 2 . » 

may be solved have been in existence for at least 50 years, and 
many types are described in the literature. The more desirable 
machines have the fewest parts per equation to be solved and 
therefore have fewest sources of error and are least expensive. 

A device has been described by Wilbur 51 in which each equation 
is represented by a tape. To study the device, consider the two 
equations : 

anxi + a 12 x 2 = h 

CL2lXl + «22#2 = i>2 

A lever, having a movable roller attached to the lever arm, is 
provided for each term in the left-hand side of the equation. The 
levers are arranged in a plane in the same relative position as the 
terms in the equation. The tapes are constrained by fixed rollers 



ANALOG COMPUTING SYSTEMS 



243 



and are also passed over each movable roller which corresponds 
to a term in the equation which the tape represents (Fig. 11-3). 
The arrangement is such that as each lever arm is moved up, the 
tape is taken up by an amount proportional to the product 
of the displacement of the roller from the pivot point on the 
lever arm and the vertical displacement of the lever arm from 
its normal position. In order to solve the equation given 
as an example, the displacement of each roller from the pivot 
point of the lever arm is made proportional to the coefficients 
On, au, a,2i, and a 22 . If the x\ lever arms are displaced from 
their normal positions through an angle whose sine is propor- 
tional to Xi, and the x 2 lever arms are displaced through an angle 



21 



22 



Fig. 11-4. Mallock's circuit. 



whose sine is proportional to x 2 , the tape corresponding to the 
first equation is shortened by an amount proportional to auXi 
+ 0,12X2. Likewise, the other tape is shortened by an amount 
proportional to a 2 i£i + 0,22X2. If the Xi and x 2 lever arms are 
displaced so that the first tape is shortened by an amount 61 
and the second tape is shortened by an amount 6 2 , the x displace- 
ments are proportional to X\ and x 2 , the solution of the equations. 

A machine was designed in England in the early 1930's by 
R. R. M. Mallock 48 which uses alternating voltages and currents 
and consists of a set of adjustable transformers. This machine 
is quite elaborate and expensive, partly because of errors in the 
transformers which must be compensated. 

If a scale model of the Mallock machine were applied to the 
equation of the previous example, three transformers, each having 
a pair of adjustable coils, would be required (Fig. 11-4). 

Let 0i, the total flux in the first transformer, be allowed to 
represent xi, and </> 2 , the total flux in the second transformer, 



244 COMPUTING SYSTEMS 

be allowed to represent x 2 . Then if the number of turns in the 
coil 11 is proportional to the coefficient an, the potential induced 
in coil 11 is approximately a n xi. Similarly, the potentials 
induced in the other coils are approximately CL12X2, 0,21X1, and 
022X2, respectively. 

The flux 4>q and the turns on the third transformer are adjusted 
so that the terminal voltages of the two coils are 61 and b 2 , 
respectively. If <£ is supplied from an external source and if 
the three coils in each row are connected in a series loop, the 
fluxes <f>i and <£ 2 will adjust themselves so they are in fact pro- 
portional to Xi and x 2 . 

Precautions must be taken to see that these fluxes do not 
become excessive in the case of equations with solutions which are 
large numbers. 

Although the Wilbur and Mallock devices are rather com- 
plicated, they do converge on the solutions of suitable equations 
automatically. A more recent machine of this type has been 
developed at RCA and is described by Goldberg and Brown. 46 
The RCA device employs alternating currents, potentiometer- 
type multipliers, and Kirchhoff adders. Stabilized amplifiers 
make up for the circuit losses and allow favorable impedance 
matching. The unknowns are represented by a-c voltages, and 
these are evaluated by comparing them with the signal supply 
voltage in a bridge circuit. This technique requires a minimum 
of precision equipment. 

Other equation solvers have also been devised using servo- 
mechanisms to drive the multiplying potentiometers, but this 
seems to be an unnecessary degree of complexity. 39 

A less expensive device of the adjuster type has been described 
by Berry et al. 40 and is now commercially available from the 
Consolidated Engineering Company. Alternating currents are 
used for convenience, and potentiometer multipliers are employed. 
As the name implies, solutions are reached by repeated adjust- 
ments of potentiometer shafts in order to carry out the iterative 
procedure of Gauss and Seidel. Fairly rapid convergence can be 
realized if the equations can be arranged so that the largest 
coefficients lie on a diagonal through the center of the deter- 
minant of the system of equations. This is a characteristic of 
the type of equation for which the machine was designed. The 
commercial model will handle 12 equations and is reputed to 



ANALOG COMPUTING SYSTEMS 245 

allow solution of simultaneous equations in from a third to a 
quarter the time required on a desk calculator. 

11-7-2. Mechanical Synthesizers and Harmonic Analyzers. 
It can be shown that the synthesis of a function from its har- 
monic components and the analysis of an arbitrary function into 
its harmonic components may both proceed on the basis of 
equations of the type: 49 



S p = \d q cos (pQ^) 



Several machines which can perform this function have been 
designed and built in this country and are described in the 
literature. 42 - 47 

A machine which is a modification of the mechanical synthe- 
sizer is called the Multiharmoni graph, and its use has been 
described by Brown and Wheeler. 52 This device will plot a sum 
of harmonic terms against another sum of harmonic terms. A 
pair of nonlinear simultaneous equations which have been con- 
verted to polar form can also be solved on this machine. 

11-8. Applications of Analog Computers 

An exhaustive discussion of the applications of analog com- 
puters is beyond the scope of this survey, but a glance at the 
reference list appended 52-71 will give an idea of the large variety 
of applications which have already been examined. It has 
been found that new and unpredicted uses are found for existing 
machines as more workers learn of their characteristics. 

References 65 and 67 contain extensive bibliographies on 
applications of differential analyzers and network analyzers. 

REFERENCES 

Components and Characteristics of Differential Analyzers 

1. Amble, O., On a Principle of Connexion of Bush Integrators, Journal of 
Scientific Instruments, Vol. 23, pp. 284-287 (1946). 

2. Beard, R. E., The Construction of a Small-Scale Differential Analyzer 
and Its Application to the Calculation of Actuarial Functions, Journal 
of the Institute of Actuaries, Vol. 71, pp. 193-227 (1942). 

3. Brown, G. S., and A. C. Hall, Dynamic Behavior and Design of Servo- 
mechanisms, Transactions of the American Society of Mechanical Engi- 
neers, Vol. 68, pp. 503-524 (1946). 



246 COMPUTING SYSTEMS 

4. Brown, G. S., and D. P. Campbell, Principles of Servomechanisms (John 
Wiley & Sons, Inc., New York, 1948). 

5. Bush, V., F. D. Gage, and H. R. Stewart, A Continuous Integraph, 
Journal of the Franklin Institute, Vol. 203, pp. 63-84 (1927). 

6. Bush, V., and H. L. Hazen, Integraph Solution of Differential Equa- 
tions, Journal of the Franklin Institute, Vol. 204, pp. 575-615 (1927). 

7. Bush, V., Operational Circuit Analysis (John Wiley & Sons, Inc., New 
York, 1929). 

8. Bush, V., The Differential Analyzer. A New Machine for Solving 
Differential Equations, Journal of the Franklin Institute, Vol. 212, pp. 
447-488 (1931). 

9. Bush, V., and S. H. Caldwell, A New Type of Differential Analyzer, Jour- 
nal of the Franklin Institute, Vol. 240, No. 4, pp. 255-326 (October, 1945). 

10. Gardner, M. F., and J. L. Barnes, Transients in Linear Systems (John 
Wiley & Sons, Inc., New York, 1942). 

11. Greenwood, I. A., Jr., J. V. Holdam, Jr., and D. MacRae, Jr., Electronic 
Instruments, Radiation Laboratory Series, Vol. 21 (McGraw-Hill Book 
Company, Inc., New York, 1948). 

12. Harris, H., The Analysis and Design of Servomechanisms, Report 454, 
Office of Scientific Research and Development, Washington, D.C. 
(January, 1942). 

13. Hartree, D. R., The Differential Analyzer, Nature, Vol. 135, pp. 940- 
943 (June 8, 1935). 

14. Hartree, D. R., and A. Porter, The Construction of a Model Differential 
Analyzer, Memoirs and Proceedings of the Manchester Literary and 
Philosophical Society, Vol. 79, pp. 51-72 (July, 1935). 

15. Hartree, D. R., The Mechanical Integration of Differential Equations, 
Mathematical Gazette, Vol. 22, pp. 342-363 (1938). 

16. Hartree, D. R., Recent Developments in Calculating Machines, Journal 
of Scientific Instruments, Vol. 24, pp. 172-176 (July, 1947). 

17. Hazen, H. L., Theory of Servomechanisms, Journal of the Franklin 
Institute, Vol. 218, No. 3, pp. 279-330 (September, 1934). 

18. Hazen, H. L., J. J. Jaeger, and G. S. Brown, An Automatic Curve Fol- 
lower, Review of Scientific Instruments, Vol. 7, pp. 353-357 (September, 
1936). 

19. Jackson, J. M., Solution of Differential Equations by an Electronic Differ- 
ential Analyzer, Admiralty Computing Service Report (Department of 
Scientific Research and Development, London; reprinted by the Office 
of Research and Inventions, Navy Department, Washington, D.C, 
July, 1946). 

20. James, H. M., N. B. Nichols, and R. S. Phillips, Theory of Servomechan- 
isms, Radiation Laboratory Series, Vol. 25 (McGraw-Hill Book Company, 
Inc., New York, 1947). 

21. Koehler, J. S., An Electronic Differential Analyzer, Journal of Applied 
Physics, Vol. 19, pp. 148-155 (1948). 

22. Kuehne, H. P., and H. A. Peterson, A New Differential Analyzer, 
Transactions of the American Institute of Electrical Engineers, Vol. 63, 
pp. 221-228 (May, 1944). 



ANALOG COMPUTING SYSTEMS 247 

23. Lennard-Jones, J. E., M. V. Wilkes, and J. B. Bratt, The Design of a 
Small Differential Analyzer, Proceedings of the Cambridge Philosophical 
Society, Vol. 35 (III), pp. 485-493 (July, 1939). 

24. Linville, T. M., and J. S. Woodward, Selsyn Instruments for Position 
Systems, Electrical Engineering, Vol. 53, p. 953 (1934). 

25. Massey, H. S. W., J. Wylie, and R. A. Buckingham, A Small Scale 
Differential Analyzer: Its Construction and Operation, Proceedings of 
the Royal Irish Academy, Vol. 45, pp. 1-21 (1938). 

26. Murray, Francis J., The Theory of Mathematical Machines, 2d rev. ed. 
(King's Crown Press, New York, 1948). 

27. Princeton University, Bibliography on Feedback Amplifiers, Technical 
Report 5 under Contract N6onr-105 Task Order VI between the Office 
of Naval Research, Navy Department, and Department of Electrical 
Engineering, Princeton University, Princeton, N.J. (March, 1948). 

28. Princeton University, Circuits for a Linear Negative Resistance and for a 
Constant Voltage-Ratio Power Amplifier, Technical Report 2 under 
Contract N6onr-105 Task Order VI between the Office of Naval 
Research, Navy Department, and Department of Electrical Engineer- 
ing, Princeton University, Princeton, N.J. (December, 1947). 

29. Sack, H. S., R. T. Beyer, G. H. Miller, and J. W. Trischka, Special 
Magnetic Amplifiers and Their Use in Computing Circuits, Proceedings 
of the Institute of Radio Engineers, Vol. 35, pp. 1375-1382 (November, 
1947). 

30. Shannon, C. E., Mathematical Theory of the Differential Analyzer, 
Journal of Mathematics and Physics (Massachusetts Institute of Technol- 
ogy), Vol. 20, pp. 337-354 (1941). 

31. Stibitz, G. R., A New Class of Computing Aids, Mathematical Tables 
and Other Aids to Computation, Vol. Ill, No. 23, pp. 217-221 (July, 
1948); Function Unit Theory of Computing Devices, Report 1309 (May, 
1948). Available on request from Taller and Cooper, Inc., Brooklyn, 
N.Y. 

32. University of California at Los Angeles, Staff of the Department of 
Engineering, The Differential Analyzer of the University of California 
(1947). Pamphlet published by the University of California, Los 
Angeles. 25 pp. 

33. Zagor, H., Description of a Reeves Electronic Analog Computer 
(REAC), Mathematical Tables and Other Aids to Computation, Vol. Ill, 
No. 24, pp. 326-327 (October, 1948). Notes on paper presented by 
Mr. Zagor before the American Physical Society, Apr. 30, 1948. 

Network Analyzers 

34. Hazen, H. L., H. R. Schurig, and M. F. Gardner, The MIT Network 
Analyzer, Transactions of the American Institute of Electrical Engineers, 
Vol. 49, p. 1102 (July, 1930). 

35. Kuehne, H. P., and R. G. Lorraine, A New A-C Network Analyzer, 
Transactions of the American Institute of Electrical Engineers, Vol. 57, 
pp. 67-73 (February, 1938). 

36. Peterson, H. A., and C. Concordia, Analyzers for Use in Engineering 



248 COMPUTING SYSTEMS 

and Scientific Problems, General Electric Review, Vol. 49, pp. 29-37 
(September, 1945). 

37. Princeton University, A Network for Solving Certain Ordinary Non- 
linear Differential Equations by Analogy, Technical Re-port 1 under Con- 
tract N6ori-105 Task Order VI between the Office of Naval Research, 
Navy Department, and Department of Electrical Engineering, Prince- 
ton University, Princeton, N.J. (October, 1947). 

38. Spangenburg, K., and G. Walters, An Electrical Network for the Study of 
Electromagnetic Fields, Technical Report 1, ONR Contract N6-ori-106 
Task III (May 15, 1947). 

Special Analog Systems 

39. Adcock, W. A., An Automatic Simultaneous Equation Computer and 
Its Use in Solving Secular Equations, Review of Scientific Instruments, 
Vol. 19, pp. 181-187 (March, 1948). 

40. Berry, C. E., D. E. Wilcox, S. M. Rock, and H. W. Washburn, A Com- 
puter for Solving Linear Simultaneous Equations, Journal of Applied 
Physics, Vol. 17, No. 4, pp. 262-272 (May, 1946). 

41. Booth, A. D., Two Calculating Machines for X-ray Crystal Structure 
Analysis, Journal of Applied Physics, Vol. 18, pp. 664-666 (July, 1947). 

42. Brown, S. L., A Mechanical Harmonic Synthesizer-Analyzer, Journal 
of the Franklin Institute, Vol. 228, pp. 675-694 (1939). 

43. Dietzold, R. L., The Isograph, A Mechanical Root Finder, Bell Labora- 
tories Record, Vol. 16, pp. 130-134 (1937). 

44. Frame, J. S., Machines for Solving Algebraic Equations, Mathematical 
Tables and Other Aids to Computation, Vol. 1, No. 9, pp. 337-351 (Jan- 
uary, 1945). 

45. Frost, A. A., and M. Tamres, A Potentiometric Secular Equation Com- 
puter, Journal of Chemical Physics r Vol. 15, pp. 383-390 (June, 1947). 

46. Goldberg, E. A., and G. W. Brown, An Electronic Simultaneous Equa- 
tion Solver, Journal of Applied Physics, Vol. 19, pp. 338-345 (April, 
1948). 

47. Hagg, G., and T. Laurent, A Machine for the Summation of Fourier 
Series, Journal of Scientific Instruments, Vol. 23, pp. 155-158 (1946). 

48. Mallock, R. R. M., An Electrical Calculating Machine, Proceedings of 
the Royal Society, Vol. 140, p. 457 (1933). 

49. Murray, F. J., The Theory of Mathematical Machines (King's Crown 
Press, New York 1948). 

50. Ragazzini, J. R., Robert H. Randall, and F. A. Russel, Analysis of 
Problems in Dynamics by Electronic Circuits, Proceedings of the Insti- 
tute of Radio Engineers, Vol. 35, pp. 444-452 (May, 1947). 

51. Wilbur, J. B., The Mechanical Solution of Simultaneous Equations, 
Journal of the Franklin Institute, Vol. 222, pp. 715-724 (1936). 

Analog Computer Applications 

52. Brown, S. L., and L. L. Wheeler, Use of the Mechanical Multiharmoni- 
graph for Graphing Types of Functions and for Solution of Pairs of 



ANALOG COMPUTING SYSTEMS 249 

Nonlinear Simultaneous Equations, Review of Scientific Instruments, 
Vol. 13, pp. 493-495 (1942). 

53. Brown, S. L., and L. L. Wheeler, The Use of a Mechanical Synthesizer 
to Solve Trigonometric and Certain Types of Transcendental Equations, 
and for the Double Summations Involved in Patterson Contours, 
Journal of Applied Physics, Vol. 14, pp. 30-36 (1943). 

54. Bush, V., and S. H. Caldwell, Thomas-Fermi Equation Solution by the 
Differential Analyzer, Physical Review, Vol. 38, pp. 1898-1902 (1931). 

55. Carter, G. K., Numerical and Network-analyzer Tests of an Equivalent 
Circuit for the Elastic Field, Journal of Applied Mechanics, p. A- 162 
(September, 1944). 

56. Carter, G. K., and G. Kron, A-C Network Analyzer Study of the Schro- 
dinger Equation, Physical Review, Vol. 67, pp. 44-49 (January, 1945). 

57. Clarke, E., and S. B. Crary, Stability Limitations of Long-distance A-C 
Power Transmission Systems, Transactions of the American Institute of 
Electrical Engineers, Vol. 60, pp. 1051-1059 (1941). 

58. Concordia, C, Network and Differential Analyzer Solutions of Tor- 
sional Oscillation Problems Involving Nonlinear Springs, Journal of 
Applied Mechanics, Vol. 12, pp. A43-A47 (March, 1945). 

59. dePackh, D. C, A Resistor Network for the Approximate Solution of 
the Laplace Equation, Review of Scientific Instruments, Vol. 18, pp. 
798-799 (1947). 

60. Hartree, D. R., and J. R. Womersley, A Method for the Numerical 
or Mechanical Solution of Certain Types of Partial Differential Equa- 
tions, Proceedings of the Royal Society (London), Vol. 161A, pp. 353- 
366 (1937). 

61. Hartree, D. R., and A. Porter, The Application of the Differential 
Analyzer to Transients on a Distortionless Transmission Line, Journal 
of the Institution of Electrical Engineers (London), Vol. 83, pp. 648-656 
(1936). 

62. Hartree, D. R., The Application of the Differential Analyzer to the 
Evaluation of Solutions of Partial Differential Equations, Proceedings 
of the First Canadian Mathematical Congress, Montreal, 1945 (University 
of Toronto Press, Toronto, 1946), pp. 327-337. 

63. Kron, G., and G. K. Carter, Numerical and Network Analyzer Tests 
of an Equivalent Circuit for Compressible Fluid Flow, Journal of the 
Aeronautical Sciences, Vol. 12, p. 232 (April, 1945). 

64. Levin, J. H., On the Approximate Solution of a Partial Differential 
Equation on the Differential Analyzer, Mathematical Tables and Other 
Aids to Computation, Vol. Ill, No. 23, pp. 208-209 (July, 1948). 

65. Maginniss, F. J., Differential Analyzer Applications, General Electric 
Review, Vol. 48, pp. 54-59 (May, 1945). 

66. Massachusetts Institute of Technology, Department of Electrical Engi- 
neering, Center of Analysis, Tables of Supersonic Flow around Yawing 
Cones, Report 3 under Contract NOrd 9169. 

67. Peterson, H. A., and C. Concordia, Analyzers for Use in Engineering and 
Scientific Problems, General Electric Review, Vol. 48, No. 9, pp. 29-37 
(September, 1945). 



250 COMPUTING SYSTEMS 

68. Reed Research, Inc., A Constant Power Circuit and Its Application to 
Instrumentation and Computation; Unpublished, Washington, D. C; 
Oct. 12, 1949. 

69. Schumann, T. E. W., The Principles of a Mechanical Method for Cal- 
culating Regression Equations and Multiple Correlation Coefficients 
and for the Solution of Simultaneous Linear Equations, Philosophical 
Magazine, Vol. 29, Seventh Series, pp. 258-273 (March, 1940). 

70. Walters, G., and R. L. Tanner, Universal Tuning Curves for Reentrant 
Cavity Resonators, Technical Report 2, ONR Contract N6-ori-106 Task 
III (Aug. 15, 1947). 

71. Whinnery, J. R., C. Concordia, W. Ridgeway, and G. Kron, Network 
Analyzer Studies of Electromagnetic Cavity Resonators, Proceedings 
of the Institute of Radio Engineers, Vol. 32, pp. 360-367 (June, 1944). 



CHAPTER 12 
THE FORM OF A DIGITAL COMPUTER 

12-1. Introduction 

In the chapters which have preceded this one, the means for 
performing automatic calculations have been described in general 
terms. In particular, a detailed description of elementary com- 
ponents has been given, and it has been shown how these com- 
ponents can be combined to perform arithmetic operations 
rapidly. Also a description of some necessary steps in the 
integration of components into a machine has been presented. 
Since these machines depend essentially on the system of arith- 
metic which is used, and since some systems are more efficiently 
utilized in computing equipment than others, a discussion of 
arithmetic systems was included. Since the form of a calculation 
depends on the details of the means chosen to carry out various 
simple arithmetical steps (such as extraction of a square root), 
generalities concerning the arithmetic components which may 
be used as steps in a calculation have been discussed. In order 
that the size of a calculation at hand may be estimated in terms 
of machinery involved, desk calculators, punched-card computing 
systems, existing and planned digital computers, and existing 
analog computers have been described; these descriptions were 
incomplete, for they were intended only to give some idea of the 
relative extents of the calculations which are suitably carried out 
on the various machines. 

In the chapters which follow, the circuits and the mechanisms 
of the physical computing and data-handling components which 
are now available and which are expected shortly to be available 
are described in sufficient detail to permit rough estimates of 
applicability and the expense of applying them in machines. 
Chapter 12 is intended as an aid in choosing components. 

12-2. The Field of Application 

When a large-scale digital computing machine is to be built, a 
field of application for this machine presumably exists and is 

251 



252 COMPUTING SYSTEMS 

fairly well defined. As a matter of practicability and economy, 
the machine may be designed to carry out the calculations 
involved in problems in this field only. 

The designation of a machine as a general-purpose machine 
simply indicates that the field from which problems are likely 
to be extracted is reasonably broad. A special-purpose machine 
is a machine designed to solve problems from a narrower field. 
However, one can expect to discover that any given machine, 
even a machine designed for a rather narrow class of problems, 
can be used effectively to solve problems outside the class for 
which it was designed. In short, it is practically impossible to 
design a machine so specialized that it will have value only with 
respect to the field of application originally intended. Nor is 
there any computer which is superior to any other computer 
with respect to every problem. In short, there is no completely 
general-purpose machine. 

The problem of designing a machine is then defined only 
approximately by the field of calculations to which it is to be 
applied. In general, the narrower the field, the simpler or the 
faster the machine. However, there may be some question 
whether a net gain is achieved by building a special-purpose 
machine. A conventional general-purpose computer may be 
capable of sufficient speed, and it may be less expensive to 
acquire such a machine than to develop a more highly specialized 
though possibly simpler computer. 

The final test of the applicability of a machine to a field of 
problems lies in an explicit demonstration of a program under 
which the machine will carry out each of the calculations required. 
From this program and from the machine's speed and reliability, 
an estimate of the speed and reliability of the calculation can be 
made. There is no other way to make this objective estimate. 

12-3. The Estimate of Size and Speed 

Any calculation must be factored into a set of elementary arith- 
metic operations before it can be carried out. Once this factoriza- 
tion has been accomplished, it is possible to estimate the size of 
the calculation. In Chap. 5 this factorization of a simple 
problem was illustrated, and an explicit program for the particular 
problem was set forth. If a time limit is placed on the length of 
the calculation, either arbitrarily or naturally through the 



THE FORM OF A DIGITAL COMPUTER 253 

application, then the rate at which the various arithmetic opera- 
tions must be carried out is determined, and from this rate the 
size of the machine required can be estimated. In particular, 
under an assumed rate of multiplication, for example, the number 
of multiplying units required can be determined. If this number 
exceeds one, the machine is likely to be so complex that its 
application will be limited, and its operation as a general-purpose 
computer will be difficult. It will then be necessary to examine 
the factorization of the calculation program critically to deter- 
mine how more than one unit can be utilized. Thus, if each 
multiplication depends for one factor upon all preceding multi- 
plications, there is no point in increasing the number of multiply- 
ing units, for only one multiplication can be carried out at one 
time in any event. On the other hand, if the problem consists 
of nothing but matrix operations on vectors, one multiplying 
unit might be conveniently utilized to compute each of the com- 
ponents of a vector. 

In general, the extent of the complexities associated with 
multiunit computing machines is great. The chief motivation 
for increasing the speed of computing units has been the desire 
to reduce the number of units in a machine without reducing its 
speed. 

Specifically, the choice of the components built into a machine 
must be based upon the programming of representative problems 
from the field to be attacked. This programming Avill indicate 
the amount of storage required, the nature of the input required, 
the nature of the output, the rate of calculation in terms of 
elementary operations of the various kinds per second, and the 
number and nature of the different operations the machine must 
carry out. From these numbers an estimate of the feasibility 
of the machine can be obtained by a search through the lists of 
components; if no components exist which provide the facilities 
required, either a more efficient reprogramming is required or, 
failing this, a more modest goal must be set for immediate 
attainment. 

12-3-1. An Example. Suppose it is desired to calculate, by 
means of a digital computer, the position of a target tracked by 
Raydist 2 equipment. Consideration of this problem will provide 
us with an example of the factoring of a calculation into a set of 
elementary arithmetic operations for machine solution. 



254 COMPUTING SYSTEMS 

Raydist is a radio navigation and tracking system. It is 
based on the measurement of the relative phase relationship 
between continuous-wave signals received at different locations. 

When the Raydist system is applied to the problem of tracking 
a target through three-dimensional space, four fixed stations are 
required. These may be represented in cartesian coordinates as 

Pi{x h y h zi) ; P2(x 2 ,y2,z 2 ) ; P*(x Zi y 3 ,z z ) ; Pi(x 4 ,y 4 ,z 4 ) 

The position of the target is P{x,y,z). Comparison circuits 
in the Raydist equipment give values Vi, v 2 , and v 3 , which are the 
differences between the distance from the target to a fixed reference 
point Pa and the distance from the target to each of Pi, P 2 , 
and P 3 . v h v 2 , and v z are related to the coordinates of the four 
fixed points and the target by the three quadratic polynomial 
equations in three unknowns which follow : 



vi = V(x — z 4 ) 2 + (y - y 4 ) 2 + (z - z 4 ) 2 

) 2 + (z - z,) 2 

(12-1) 



- VO - zi) 2 + {y - 2/1) 2 + (z - zi)' 
v 2 = y/{x — x A ) 2 + (y — 2/ 4 ) 2 + (z — Zi) 2 

- V(% - Z2) 2 + (y - 2/ 2 ) 2 + (z - z 2 )' 



vz = V(x - X4) 2 -f- (y - y A ) 2 + (z - z 4 ) 2 

- V(x - xz) 2 + (y - 2/ 3 ) 2 + (z - z 3 ) 2 

These equations are equations of hyperboloids of revolution, 
each with one focus at the point P 4 . The other foci are at points 
Pi, P 2 , and P 3 , respectively. The problem is to solve for the 
coordinates (x,y,z) of P, knowing all the other numbers. 

A Formulation of the Problem. One formulation of this 
problem is based on the assumption that the target moves with 
restricted accelerations and that a fairly continuous record of its 
position is desired. Thus, a system of sharpening an estimated 
path of the target will be adequate. The position after any 
step can be estimated on the basis of its position after the preced- 
ing step and its average velocity during the preceding step. 
Thus, if the calculations of position P occur after equal incre- 
ments of time, so that the kih calculation occurs at the time k At 
with At constant, and if p k (x k ,y k ,z k ) denotes the position at the kth 
time interval, then the coordinates of P k+1 can be estimated as 
x k+i = 2 x k — x k ~ l 

y k+l = 2 yk - yk-1 (12-2) 

z k+1 = 2z k - z k ~ l 



THE FORM OF A DIGITAL COMPUTER 255 

The accuracy of this estimate is guaranteed by the size of At and 
the extent to which the acceleration is restricted. It is to be 
assumed that the initial position P° is known accurately. 

With this formulation an extension of Newton's method may be 
applied. For this example Eqs. (12-1) are reduced to polynomial 
form: 

AiX 2 + Biy 2 + dz 2 - 2Diyz - 2EiZX - 2Fay - 2G { x 

- 2H iV - 2Uz = V i} i = 1, 2, 3 (12-3) 

Here all the coefficients are functions of the known numbers in 
Eqs. (12-1) (the coordinates of Pi, P 2 , P 3 , Pi and the numbers v h 
V2, v 3 ). The algebra involved in the transition from Eqs. (12-1) 
to Eqs. (12-3) consists in rearranging Eqs. (12-1) so that a radical 
is isolated on one side of the equation, squaring, again rearranging 
so that the only remaining radical is isolated, and again squaring. 
The process can be carried out using the literal coefficients shown 
in Eqs. (12-1), and explicit formulas for the coefficients in Eqs. 
(12-3) in terms of these letters can be exhibited. 

Once the coefficients in Eqs. (12-3) are known, the approximate 
values of x, y, and z to be sharpened can be substituted in the 
left members. For this substitution the amounts that Vi, V 2 , 
and V 3 differ from the actual values generated by the approximate 
x, y, and z may be denoted by AVi, AV 2 , AV 3 ; the problem is to 
modify x, y, z by changing to values x + Ax, y + Ay, z + Az so 
that changes of the left-hand members of the Eqs. (12-3) induced 
by the change in x, y, z are just AVi, AV 2 , AVz. This can be 
achieved approximately by linearizing the problem through the 
use of differentials and solving the three equations: 

d Jl dx + <!p d y + ?p dz = dV i = 1, 2, 3 (12-4) 

ox oy oz 

simultaneously for dx, dy, and dz in terms of dVi, dV 2 , dVs. If 
the values AFi, AV 2 , AV 3 are substituted in place of dVi, dV 2 , dV 3 , 
numbers corresponding to dx, dy, dz may be found. These 
differentials are approximately equal to the actual Ax, Ay, Az, 
which must be added to the assumed values of x, y, z in order to 
account for accelerations during the time interval A£. Thus, 
the approximate solution will have been sharpened by the method 
of Newton. 

In summary, this approach to the solution involves the fol- 
lowing steps : 



256 COMPUTING SYSTEMS 

1. Read a h a 2 , and a 3 from Raydist dials. (These are the 
phase differences between signals as received at different 
stations.) Multiply by a factor k to convert dial readings 
from wavelength of radio-frequency carrier to distances v h 
v 2 , v 3 in some convenient scale. 

2. From Vi computed in step 1 and constants Li, Mi, . . . , T% 
involving the coordinates of Pi, P 2 , Pz, Pa, compute A { , B i} 
d, etc., the coefficients of Eqs. (12-3). 

3. From the values of x k ~ 1 , y k ~ 1 , z k ~ l , x k , y k , and z k computed 
in the two previous calculations, and with Eqs. (12-2), 
calculate approximate values of x k+1 , y k+l , and z k+l . 

4. Use the coefficients calculated in 2 and the approximate 
distances calculated in 3 to compute the nine partial-deriva- 
tive coefficients of Eqs. (12-4). 

5. Assemble the coefficients found in 4 into a Jacobian matrix 
and compute the value of the nine minor determinants of 
the nine matrix terms. (Note: The Jacobian matrix of a 
set of differential equations is the matrix formed by the 
derivative coefficients.) 

6. From the terms computed in 4 and the minors computed in 
5, compute the value of the Jacobian determinant of the 
Eqs. (12-4). 

7. Substitute the coefficients of Eqs. (12-3) computed in 2 
and the approximate distances computed in 3 into the left- 
hand sides of Eqs. (12-3). Subtract the value computed 
from the values of Vi computed for the right-hand sides of 
the equations. Label the differences AV{. 

8. Use the values of AV* computed in 7 as substitutes for dVi, 
and the values of the minors and the determinant computed 
in 5 and 6 to solve Eqs. (12-4) for dx, dy, and dz. Label the 
results Ax, Ay, and Az, respectively. 

9. Add Ax, Ay, and Az to x k+l , y k+1 , and z k+l , respectively. 
These sums are the sharpened values of x k+l , y k+1 , and 
z k+1 . 

Breakdown of the Problem into Elementary Steps. In order to 
determine the size and speed of the computer required to handle 
the problem just outlined in a given length of time, it is necessary 
first to arrange the steps in a chronological order which will 
allow the computation to proceed directly from the input data 



THE FORM OF A DIGITAL COMPUTER 257 

to the final result. This has been done in the foregoing section. 
It is necessary further to count the operations of each type. An 
illustration of how this might be done is presented in tabular form 
in Table 12-1, along with a symbolic representation of the opera- 
tions listed in the preceding section. 

Table 12-1 shows the number and nature of the various opera- 
tions required to estimate coordinates and to sharpen the estimate 
once. Slight deviations can be made from the schedule outlined, 
and in a binarj^ computer, multiplication by powers of 2 may be 
accomplished simply by shifting the multiplicand. Since 
division is required only once, it might be desirable to perform 
division by an iterative process involving only addition and 
multiplication rather than provide a divider unit for this par- 
ticular problem. A description of such a method appears in the 
Annals of the Commutation Laboratory of Harvard University 1 
and is repeated on pages 122-123 supra. 

The time interval At is established on the basis of the maneu- 
verability or possible acceleration of the target and the precision 
with which it is desired to track the target. The latter factor 
also determines the size of the numbers to be handled in the 
computer. For purposes of illustration, let it be assumed that 
the coordinates must be calculated every 0.1 second in order to 
render valid the approximations involved in the method outlined. 

The requirement for 10 solutions per second demands that 
automatic input equipment be employed. The detailed specifica- 
tions for automatic input equipment depend upon the output 
of the measuring equipment and the arithmetic system used in 
the computer. The primary physical quantities representing 
the output of the Raydist system are a pair of audio-frequency 
voltages which differ in phase and actuate an integrating syn- 
chroindicator. Each cycle of phase shift might be counted with 
a set of electronic counters such as those described in Sec. 
13-5. Undoubtedly other devices would evolve from a careful 
study of the problem. Whether or not the digital data from the 
input equipment are fed directly into the arithmetic units of the 
computer or into a storage device depends upon whether on-line 
or off-line operation is being sought. In the former case, the 
computer solves for points on the trajectory as the experiment is 
performed ; in the later case, the computations may be performed 
at a later time and at a scaled-down rate. 



258 



COMPUTING SYSTEMS 
Table 12-1 













Arithmetic processes 


Step 
No. 


Operation 


Add 


Multi- ~. 
i )i- 






or sub- 


ply or 


vide 






tract 


square 




1 


Vi = kai 




3 




2 


4»? = 4(^) 2 

Ai = 4»J + ^ 

Bi = \v\ + Mi 

d = 4v\ + #»• 

Di, Ei, and F t - are constants. 


3 
3 
3 


6 






d = 4Miv] + Ni 


3 


3 






Hi = Wivl + Pi 


3 


3 






Ii = 4Qivl + Rt 


3 


3 






vt = {vlY 




3 






Vi = v\ + SiV 2 + Ti 


6 


3 




3 


Approximate values: 










4 +1 = 2x k - x*- 1 


1 


1 






y k+l = 2 yk _ yk-l 


1 


1 






z\ +l = 2z k - z k ~ l 


1 


1 




4 


^ = 2(Atx - EiZ - F iV - Gt) 
ox 


9 


12 






^ = 2(B iy - DiZ - Fix - Hi) 
dy 


9 


12 






?p = 2{CiZ - D iV - E iX - h) 

oZ 


9 


12 




5 


Let 










dVi dV 2 dVz 












dx dx dx 












dVx dV 2 dVz 


= Dj, determinant of Ja- 










dy dy dy 


cobian matrix 










dVx dV 2 dVz 












dz dz dz 












Find the nine minors, where 












dV 2 dVz 












M 1<x = 


dy dy 

dVj dV 3 

dz dz 




9 


18 




6 


dVi „, dVx ,- , dVi , T 


2 


3 






Find x 2 , y 2 , z 2 , 2yz, 2zx, 2xy, 2x, 2y, 2z 




12 



THE FORM OF A DIGITAL COMPUTER 
Table 12-1. (Continued) 



259 





Operation 


Arithmetic processes 


Step 
No. 


Add 
or sub- 
tract 


Multi- 
ply or 
square 


Di- 
vide 


7 


AVi = (Aix 2 + B iy * + Ciz* - 2D iV z - 2E { zx 
- 2Fixy - 2GiX - 2H iV - 2I iZ ) - Vi 


9 


27 




8 


± = 1 + Dj 

A7,(M U ) - AVziMi.y) + A7,(¥i,,) 
dX ~ Dj 

-AFi(M 2 .x) + AF 2 (M 2>J/ ) - AF 3 (M 2 , Z ) 


2 
2 
2 


4 

4 
4 


1 




rf2/ A, 

J AFi(M 3 .x) - AV 2 (M 3 , y ) + AF 3 (M 3 , 2 ) 
d2= Dj 




9 


Improved: 

x k+i ~ x k+i + dx 
yk+i ~ yk+i + ^ 
2 fc+i ~ gj+i + dz 


1 
1 
1 






T 


otals 


83 


135 


1 









In the assumed Raydist problem the chief difficulty is in per- 
forming the arithmetic operations fast enough. Approximately 
135 multiplications are required for each estimate, so if one 
estimate is adequate and only one multiplier is employed, 
multiplications must be performed in less than 700 microseconds. 
It is shown in Chap. 13 that the fastest binary multipliers now 
known to exist operate in about 5n microseconds, where n is the 
number of binary digits in the numbers to be multiplied per 
microsecond of pulse width. If a 1-microsecond pulse width 
and a 30-binary-digit precision are assumed, each multiplication 
could be carried out in 150 microseconds, which is safely within 
the required time. 

If the calculations are performed on a high-speed general- 
purpose machine, they will probably be taken sequentially. An 
examination of the steps listed in Table 12-1 reveals that the 
quantities to be stored in addition to the three incoming a,- and 
the k, Li, Mi, . . . , Ti constants of the Raydist system are 



260 COMPUTING SYSTEMS 

almost all of the 83 sums and the 135 products. Thus, the 
number of quantities to be stored is approximately 275, and it 
will be seen in Chap. 14 that this is a relatively small number for 
present-day storage systems. 

The output requirements in this example depend almost 
entirely upon uses to which the points in the trajectory are to 
be put. In some cases the position of the target transmitter 
might be used directly to control some other process. In other 
cases, a permanent tabular record of the coordinates of the 
trajectory might be desired. Still another possibility is that the 
trajectory should be traced by some plotting device for visual 
inspection. Equipment with which to perform these functions 
is described in Chap. 15. It is possible that it may be satis- 
factory to use output data taken at longer time intervals than the 
input data, since the time interval at which the input data is 
sampled was determined by the requirements of the mathe- 
matical method employed. 

12-4. Conclusions 

In the example discussed, the crux of the difficulty in perform- 
ing on-line computing with a general-purpose machine comes in 
the number of multiplications required. In other problems the 
input equipment or the storage system may prove to be the source 
of technical difficulty. In any case, a program similar to the one 
outlined for the example given in this chapter must be carried 
through in order to find the machine requirements for the various 
possible mathematical approaches. In this way, it may be found 
that some methods are better suited to the computing machinery 
available than other methods which might have seemed more 
inviting. 

After a preliminary estimate of the relative speeds and capabili- 
ties required in various parts of the machine, a complete coding 
of the problem must be made before an accurate determination 
of speed can be arrived at. 

The question of whether on-line or off-line operation is more 
suitable can be determined only by applying the speed require- 
ments to those physical components which are available. Like- 
wise the question of whether or not a general-purpose computer or 
a faster special-purpose computer is required depends ultimately 
upon whether or not the speeds of the arithmetic units available 



THE FORM OF A DIGITAL COMPUTER 261 

are great enough to effect a complete solution in the time allotted. 
In cases where on-line operation is required because of the nature 
of the problem, and where computing speeds are not sufficiently 
high, a certain amount of parallel operation is required, and the 
designer and the mathematician must devise means of accom- 
plishing various steps simultaneously. 

REFERENCES 

1. Harvard University, Staff of the Computation Laboratory, A Manual 
of Operation for the Automatic Sequence Controlled Calculator, Annals of 
the Computation Laboratory of Harvard University, Vol. I (Harvard 
University Press, Cambridge, Mass., 1946). 

2. Hastings, Charles E., Raydist — a Radio Navigation and Tracking Sys- 
tem, Tele-Tech., Vol. 6, No. 6, pp. 30-34 (June, 1947). 



Part III 
PHYSICAL COMPONENTS AND METHODS 



The following chapters include a detailed discussion of physical 
components which have been introduced in Chaps. 3 and 4 
of Part I and which have been described in Part II with regard to 
their integration into computing systems. In Chap. 13 is 
described the organization of simple arithmetic units using 
elementary potential digital computing components and switches 
and gates. Chapter 14 deals with the transfer of data in time 
and in space. Storage is defined as a transfer of data in time; 
switches and cables throughout a computer are described as 
components effecting a transfer in space. Data-conversion 
equipment, including input, output, and radix-changing devices, 
is discussed in Chap. 15. Special physical techniques of possible 
use in automatic calculation are summarized in Chap. 16. 
Finally, Chap. 17 outlines some of the factors affecting the choice 
of components. 



CHAPTER 13 
ARITHMETIC ELEMENTS 

13-1. Introduction 

Since the first Oriental abacus was constructed, many ingen- 
ious units have been devised to perform arithmetic operations. 
The arithmetic elements to be discussed here, however, are 
restricted to those of an electronic nature, in which electron tubes 
and crystal rectifiers are employed to carry out the desired opera- 
tions. These electronic elements have their counterparts in 
electromechanical relay systems, mechanical counter wheels, 
and other mechanical devices. These will not be discussed 
specifically here. However, specific examples will be given 
in order to indicate the manner in which high-speed arithmetic 
elements are built up from individual switches, gates, and 
counters. These examples will serve to indicate how other 
more complicated systems may be designed from basic com- 
ponents of any type and to give an idea of the total amount of 
equipment required to meet various specifications. The differ- 
ences between serial and parallel arithmetic operations are 
reviewed so that elements performing these operations may be 
comprehended. The characteristics of fixed-cycle and variable- 
cycle operation are also mentioned so that the differences in their 
requirements may be understood. Examples of coincidence 
adders and ring-type adders are described. Finally, a discussion 
has been included of the incorporation of adders into complete 
arithmetic units. Emphasis has been given to arithmetic units 
of the binary type, but these may be expanded to compute in 
radix higher than 2 by an extension of the procedure outlined in 
this chapter. In the material which follows, it may be assumed 
that numbers are binary unless otherwise designated. 

13-2. Series versus Parallel Operation of Arithmetic Units 

An initial choice which must be made in deciding the method 
of computation in an automatic digital computer is whether 

265 



266 



PHYSICAL COMPONENTS AND METHODS 



serial or parallel arithmetic elements are to be used. To indicate 
the fundamental difference between serial and parallel arith- 
metic units, consider two such elementary adders. In a serial- 
type adder, as shown in Fig. 13- la, the adder receives simul- 
taneously only a single pair of equal-order digits of the addend 
and augend. The lowest order, or right-hand, digits of each 
number are added first, and a sum and carry are formed. This 
carry and the second digits are then added, etc., in succession. 
Thus, if 0,638 is to be added to 0,092, the arithmetic unit first 
receives a 2 and an 8 to add. The modulo 10 sum is registered 
in the adder, and the carry digit 1 is stored to add to the pair 



ADDEND 





6 


3 


8 




ADDER 




















9 


2 


AUGEND 





SUM -0730 



CARRY - — 



ONE ADDITION TIME 



STORAGE 



Fig. 13-la. Block diagram of serial addition system. 

of second-order digits 9 and 3. These three digits are summed 
during the time called the second addition time and result in a 
sum of 3 and another carry of 1. This process is continued until 
all the sums have been formed, in this case, two more times. 
Serial arithmetic might be considered desirable in conjunction 
with a cyclic storage medium of large capacity and low access 
time. 

In a parallel adder all pairs of equal-order digits in the addend 
and augend are entered into the adder simultaneously. Thus, 
the number of adder units is equal to the number of digits in the 
largest of the numbers to be added. Each adder unit is identified 
with a digit-order number, and it is cascaded with the others in 
ascending digit order as illustrated in the block diagram of Fig. 
13-16. To form the sum of 0,638 and 0,092, of the previous 
example, the first-order digits 8 and 2 are paired together and 
entered into the right-hand adder unit. At the same time the 
second-order digits, 3 and 9, enter the second adder unit to the 
left, etc., for all four pairs of digits. Additions of all pairs 
of digits take place simultaneously. Carries are formed in 
adders I and II. These carries are passed on to the adders of 



ARITHMETIC ELEMENTS 



267 



next higher order, II and III, respectively, where they are added 
to the initial sums; this may result in further carries, so that 
in the extreme case a carry could travel from the lowest order 
right-hand adder to the highest order left-hand adder. 

The major difference between these two methods is that when 
two numbers having n significant digits are added, a parallel 
adder requires n times as many individual adder units as a serial 
adder. The serial adder, however, requires n addition time units 
in which to find the sum, whereas the parallel adder requires but 
one addition time unit plus sufficient time in which to allow n 



ADDEND 
AUGEND 





< 




[ 

9 


3 

.■r~ 


2 

1 


E 




ADDERS IV 


" 


II 


1 


SUM C 


t 


7 




3 




Ln 








Fig. 13-16. Block diagram of parallel addition system. 



successive carries. The differences between other arithmetic 
units which operate serially and in parallel are similar. 

It seems quite feasible to build computers in which both serial 
and parallel representation of numbers is employed. 5 For 
example, in a machine being built by the Raytheon Company, 
binary digits are stored serially in acoustic delay lines and are 
added in a parallel adder. Conversion of numbers from one 
form to another is fairly simple in a machine in which all pulses 
are synchronized with a master clock-pulse generator, as illus- 
trated in Figs. 13-2a and b. In each case, one-pulse time-delay 
circuits delay each digit by a number of pulse times proportional 
to the number of the flip-flop in which it is received or has origi- 
nated. Each process is started by a single pulse applied to all 
gates simultaneously. 

The shift register described in Sec. 13-6-3 may also be used as a 
serial-to-parallel converter and operates in a manner similar to 
the converters just described. 

A distinction between series and parallel operation at the 



268 



PHYSICAL COMPONENTS AND METHODS 



level of the mathematical process of addition, subtraction, and 
multiplication also exists. Although it is considered rather 
complicated to perform several of these processes simultaneously, 



SERIAL 
STORAGE 



PARALLEL 
REGISTER 



GATING PULSE — 



+m 



& 



-go 



FF-0 



FF- 



FF-2 



& 



.._rf 



Fig. 13-2a. Serial-to-parallel converter. 



FF-n 



PARALLEL 
REGISTER 



GATING PULSE — 



FF-0 



SERIAL 
STORAGE 



FF- 



m 



m 



& 




FF-n 



Fig. 13-26. Parallel-to-serial converter. 

or in parallel, in a high-speed electronic digital computer, it 
is not uncommon to use punched-card calculating machines in 
this manner. 

13-3. Operation Sequence Control 

It is possible to design a computer which requires instructions 
only to perform certain arithmetical operations in a given 
sequence, the internal control of the machine automatically 
initiating one operation upon completion of the previous opera- 
tion. In the interest of simplicity of equipment, however, most 
present-day computers require that in the preparation of the 
problem a particular time be assigned at which each operation 
shall be initiated. Since an operation such as addition may 
require more time when one set of numbers is added than when 
another set is added 3 (depending upon the number of carries 
generated) and since it is impractical in coding the problem to 



ARITHMETIC ELEMENTS 269 

predict the time requirement in advance, the maximum amount 
of time required for the greatest possible number of carries must 
be allowed. Thus, on the average, the computer which proceeds 
automatically can perform a given operation faster than one 
which must wait the maximum time for each operation. These 
two types of operation are sometimes called fixed cycle and 
variable cycle. 

As arithmetic units are improved, the time requirements for 
the two types of operation are becoming more nearly equal and 
the speed advantage of automatic control is lessened. 

13-4. Elementary Coincidence -type Adders 

Chapter 4 described an idealized adder element. This adder 
contained two counters. At the start of an addition, the addend 
was set up in one counter, called the addend register; the augend 
was set up in the other, the accumulator. The accumulator was 
then made to count. Starting from its initial position, i.e., 
starting not at zero but at the number representing the augend, 
it was made to count the addend. When the count was com- 
pleted, the accumulator contained the sum of augend and 
addend. Such an adder may be called a counter-type adder. 

A coincidence-type adder differs in that pulses representing two 
equal-order digits of addend and augend and the previous carry 
are received by the adder simultaneously, and the correct sum 
and carry pulses are gated out immediately. Its name is derived 
from the fact that the output pulses are dependent upon the 
coincidences of the input pulses. In this section a number of 
coincidence adders are described, simplified electronic circuits 
are given, and the arithmetical operations which they can per- 
form are outlined. These examples are patterned after the 
serial binary-arithmetic equipment developed during the EDVAC 
study at the University of Pennsylvania 8 and include the follow- 
ing types: (1) binary nonalgebraic adders; (2) a binary algebraic 
adder; (3) amplitude adders; (4) a coded-decimal adder; (5) 
matrix and cathode-ray-tube adders. 

13-4-1. Two-input Binary Adders. In order to perform an 
addition of two numbers in any radix system, an addend digit, 
an augend digit, and a previous-carry digit must be summed. 
These three digits may be added either as pairs in two separate 
steps or simultaneously in one step. Two types of adders have 



270 



PHYSICAL COMPONENTS AND METHODS 



been devised, one for each of these methods of adding three 
digits. 

Since the addition of only two of the three digits referred to 
constitutes something less than the entire operation, the term 
half adder has been used to signify a device in which two digits 
are added. A combination of two half adders, which can add 
three digits in two separate steps, is called a two-input adder. A 
system in which three digits are received simultaneously at the 
input is called a three-input adder. The two-input adder is 
discussed in this section. 

A half adder delivers a proper sum and carry digit for each pair 
of digits which it receives. Since these are not always the final 
sum and carry digits, they will be designated as S' and C", 
respectively. Table 13-1, a binary addition table, shows the 
sum and carry for each possible combination of two binary 
digits, A and B, and will serve to illustrate the binary addition 
process. 

Table 13-1 



Addend 


Augend 


Sum 


Carry 


A 


B 


S' 


C 

















1 


1 





1 





1 





1 


1 





1 



Suppose we want to design a device capable of performing the 
binary addition process for each of these combinations of values 
of A and B. This device, called a half adder, will consist of 
and, or, and not circuits, as described in Chap. 4, arranged so as 
to carry out the operations exemplified in Table 13-1. To deter- 
mine what this arrangement of and, or, and not components 
should be, let us state a set of rules to be used as functional 
specifications for the adder. These rules will be derived from 
an inspection of Table 13-1 and will be stated in the terminology 
of formal logic. 

It may be seen that 1 and are the two possible values of A, 
B, S', and C ', corresponding to the logical concepts true and 
false. 

The logical concepts and, or, and not, as embodied in com- 



ARITHMETIC ELEMENTS 



271 



ponents of digital computation (e.g., gating circuits), are denned 
by the following statements in the language of formal logic 4,7 
in which the letters A and B themselves denote statements which 
can be either true or false: 

1. A and B denotes the statement which is true if and only if 
both A and B are true. (Applying this definition to a 
gate circuit, A and B represents the condition which obtains 
when the value 1 appears both at A and at B.) 

2. A or B denotes the statement which is true if and only if at 
least one of the statements A, B is true. (That is, A or B 
denotes the condition which obtains in an anti-Rossi circuit 
when there is a 1 at A, at B, or at both A and B.) 

3. Not A denotes the statement contradictory to A. If not A 
is true, A is false. [That is, not A denotes the condition 
which obtains when there is a at A. Not (A and B) 
denotes the condition which obtains when the value 1 
does not appear at both A and B.] 

4. A = B denotes a statement which is true if and only if 
either A and B are true or A and B are false. (That is, 
A = B means that if there is a 1 at B, there is also a 1 at A ; 
if there is a at B, there is also a at A.) 

On the basis of these definitions, we can derive and state the 
following rules of binary addition from an inspection of Table 
13-1: 

S' = (A or B) and not (A and B) 

C" = A and B 

Figure 13-3 is a block diagram showing an arrangement of 
logical and, or, and not components which enforces these rules 
of binary addition. 



A* 


OR 




AND 




B« 












t 






NOT 




A" 


ANO 








J 




B»— 









Fig. 13-3. Half adder, block diagram. 

A vacuum-tube equivalent of the same arrangement is illus- 
trated in Fig. 13-4. 



272 



PHYSICAL COMPONENTS AND METHODS 



In Fig. 13-4 the first grids of tubes Vi and 7 2 serve the function 
of or circuits, and the second grids serve as the and-not circuits, 
since their input is received from the and circuit output of V%. 
V4 serves as a clock-pulse gate not shown in the block diagram. 
Positive pulses are required to represent 1 digits in this con- 
figuration. Vi or Yi will pass a pulse if either receives a 1 pulse 
from A or B. If a pulse enters through both A and B, T 3 passes 
a carry pulse and also cuts off Vi and Vi by driving their second 
grids negative, thereby inhibiting the pulse which would normally 




— (jQ— = CORRECTIVE DELAY 



Fig. 13-4. Half adder, schematic diagram. The initials C.P. stand for 
clock pulse. 

have come from Vi and V 2 . V ia and V±b constitute a gate 
through which clock pulses are passed when F4 is cut off. This 
is done to regenerate the pulses, as described in Chap. 4. The 
squares marked with C are corrective-delay networks, which 
are required when a tube performs a not function requiring pulse 
inhibition such as Vz performs with Vi and F 2 in Fig. 13-4. In 
order to inhibit reliably, it is necessary for the negative pulse to 
reach the second grid before the positive pulse reaches the first 
grid. This is accomplished by delaying the signals to Vi and V 2 . 
Another case where a corrective delay is employed is in the gating 
circuit which causes clock pulses to replace other pulses. In this 
case, the clock pulse should be delayed slightly. In both cases, 
the undelayed pulse should be broadened so that it overlaps the 
delayed pulse. 

The next step of the binary addition involves carry-over from 
the lower order digital addition. If a synchronous serial system 
is being employed, the carry pulses can be stored in a delay circuit 
having a one-pulse time delay. The development of the second 
half of a full binary adder stems from the rules of binary arith- 



ARITHMETIC ELEMENTS 



273 



metic illustrated by Table 13-2 in which two digits A and B 
have been added to produce S' and C", and in which the carry 
from the previous addition is C". In the first three columns are 
shown all eight possible combinations of three binary digits 
taken two at a time. The fourth and eighth rows are incom- 
plete, because a sum and carry of 1 is an impossible combination 
in binary arithmetic, just as a sum of 9 together with a carry of 1 
is an impossible combination in the decimal system when only 
two digits are added. 

Table 13-2 



Previous carry 


Initial sum 


Initial carry 


New sum 


New carry 


C" 


S' 


C" 


S 


C 























1 





1 





1 





1 








X 


I 






1 








1 





1 





1 


1 


1 


1 


1 








1 


X 


I 


I 







The rules with which this table conforms also may be stated in 
logical terms thus: 

S = (£' or C") and not (S f and C") 
C = (£' and C") or C 

These rules are similar to those of a half adder if we substitute S' 
for A and C" for B. There is an additional digit C" which must 
also be recognized by the output carry C. A full adder can be 
compounded of two half adders and an or circuit in the manner 
illustrated in Fig. 13-5. 

Figure 13-6 is a schematic diagram of a serial adder, correspond- 
ing to the block diagram shown in Fig. 13-5. Tubes Vi to V± 
form the first half adder, and V- to V% form the second. The 
squares marked C indicate the presence of clock pulses. Their 
effect is simply to synchronize the times of arrival of pulses at 
various points in the circuit. The interval between successive 
clock pulses is the interval between the instants of arrival at the 
inputs A and B of successive pairs of corresponding digits of the 
numbers to be added. 



274 



PHYSICAL COMPONENTS AND METHODS 



Suppose that two numbers A A' A" and BB'B"sLre to be added. 
A" and B" represent the least significant digits of the two num- 
bers. Suppose that each is 1. Following the notation used in 
Table 13-2, let C" designate the carry generated by their addition 
in the half adder formed by tubes V\ to F 4 . This carry pulse is 
gated through V$ and Fio into a one-pulse delay circuit repre- 



A' 


OR 




AND 


1 




OR 




AND 




e — 




-ev 




— ** SUM 










f 








t 




A 

8 


AND 


— 


— ». 


NOT 


— c^ 


AND 


^ 


NOT 
































— -C- 


CARRY TO 




l ■» 


OR 


DIGIT OF 
NEXT-*- 






HIGHER 






ORDER 



Fig. 13-5. Two-input adder, block diagram. Two half adders (see Fig. 
13-3) are combined to form this system for adding three digits in two steps. 




Fig. 13-6. Two-input adder, schematic diagram. The sum digits appear 
serially at S. Carries from each addition of S' and C" (see Fig. 13-5) are fed 
back into the second half adder through the one-pulse delay circuit, IP. 

sented by the square marked IP. Here it will be delayed for one 
clock-pulse interval. 

Now suppose that the next two digits A' and B' are 1 and 0. 
Then their sum S' is 1. The pulse representing S' arrives at F 5 
at the same instant that C" (which was delayed one pulse time) 
arrives at Ve. 

Tubes V$ to Vs constitute another half adder. Since C" and 
S' are both 1, the and tube V7 starts to conduct, blocking F 5 and 



ARITHMETIC ELEMENTS 



275 



F 6 and generating a negative pulse which is fed back into V 9 . 
V9 is an or tube. A negative pulse produced by a carry from 
either half adder will make V9 generate a carry pulse to be stored 
in the delay circuit. The pulse generated by Vi, when C" and S' 
are applied to its two grids, therefore causes Vg to store another 
carry pulse in the delay circuit. (This pulse, after being delayed, 
reaches V& simultaneously with the sum of A and B, the two 




Fig. 13-7. Two-input serial adder using only eight tube envelopes. 

digits of next higher order.) It can be shown, by following other 
possible combinations of input pulses through the circuit, that it 
will perform all the requirements specified by Table 13-2. 

It is worth noting that V9 can never be pulsed on both grids 
simultaneously because if S f is 1, C must be and there is there- 
fore no pulse on the second grid of F9; if S f is not 1, there can be 
no pulse on the first grid of V$. 

Since an adder of the type described uses nine logical units, 
and since there is usually a choice of several circuits for each 
logical unit, there are many possible variations of the simple 
coincidence adder. 

A two-input adder which is a slight variation of the one 
described is illustrated in Fig. 13-7. Only eight tube envelopes 
are used in this circuit as compared to 10 in Fig. 13-6. The 



276 



PHYSICAL COMPONENTS AND METHODS 



half adders are composed of tubes Vi to F 3 and 7 5 to V 7 , respec- 
tively. V \ is simply an inverter for V 8 , the or circuit. The out- 
put of Vs is the carry C, and it is fed into a one-pulse time-delay 
circuit, where it is stored. The output of the delay line is 
connected to the input of the second half adder, so that carry 
pulses reach the half adder at the same time as the modulo 2 
sum of the next pair of digits. 

13-4-2. Three-input Binary Adders. The second type of 
coincidence adder to be considered has three inputs and two 
outputs and may be operated either from pulses or static volt- 
ages. Thus it is more versatile than the one just discussed. 
Only one rise time is required for the carry pulse, and as a result 
less time must be allowed for a complete carry when this type 
of unit is used in a parallel adder. 

To derive the logical units of the three-input adder, consider 
the full binary addition table shown in Table 13-3. The rules 

Table 13-3 



Previous carry 


Addend 


Augend 


Sum 


Carry 


C" 


A 


B 


S 


C 























1 


1 








1 





1 








1 


1 





1 


1 








1 





1 





1 





1 


1 


1 








1 


1 


1 


1 


1 


1 



with which this table conforms are 

S = (A and B and C") or {A or B or C" and not 

[(A and B) or (A and C") or (B and C")]) 
C = (A and B) or (A and C") or (B and C") 

The block diagram shown in Fig. 13-8 illustrates a circuit 
for enforcing these rules in a computer. Many variations of 
this circuit are possible, but the following may be considered 
representative. 

Figure 13-9 is a circuit in which the digit 1 is represented by a 
positive pulse. Let us examine its behavior under each of the 



ARITHMETIC ELEMENTS 



277 



A. — 


L-j 


















8* 


OR 




AND 




OR 


»»S 




l — ' 


* 






C« — 






t 














NOT 






A» — 




AND 




































A»— 




AND 




_ 


























B« — 




AND 






C'i 








A. — 














AND 






CS— 















Fig. 13-8. Three-input adder, block diagram. 




S(SUM) 



C (CARRYJ 



Fig. 13-9. Three-input adder, schematic diagram. 



278 



PHYSICAL COMPONENTS AND METHODS 



eight conditions represented by the first three columns of the 
table. Initially the shaded tubes are conducting. If both 
F 9 and Fio are cut off, a positive pulse will appear at S. If V\ia 
is cut off, a positive pulse will appear at C. 

The first row in the table can be disposed of immediately. 
No pulses appear on any of the input leads, and no pulses are 
produced in either of the outputs S and C. 



S(SUM) 




C (CARRY) 



Fig. 13-10. Variation of three-input adder circuit. 

Each of the second, third, and fifth rows has a 1 in A or B or C" . 
Either Vi or V 2 or V 3 will be triggered; Vg and Fio will both be 
cut off; a positive pulse, or 1, will appear at S. 

The fourth, sixth, and seventh rows contain the combinations 
A and B, B and C" , and A and C" . One of these three combina- 
tions will trigger F 4 , Vi, or V&, which will prevent any of Vi, V 2 , 
and V z from conducting. There will be no pulse, i.e., 0, at S. 
At the same time, if any of V 4 , V&, or F 6 conducts, the grid of 
Vi2a will drop, the tube will be shut off, and a 1 will appear at C. 
If both A and B or both B and C" are 1, then either F 7 or Vs 
will conduct also. But either alone, cutting off only one of Vq 
and Vio, has no effect on the output S. 

The only remaining combination is illustrated in row eight, 



ARITHMETIC ELEMENTS 



279 



where all three inputs are 1. The effect of this combination 
differs from the effect of a combination of two only in that both 
Yi and Vs conduct, cutting off both F 9 and Fio, producing a 
1 at 8. 

Figure 13-10 is a slight variation of Fig. 13-9 in which tubes 
Vi, V 2 , and V 3 provide the three single inputs which can cause a 1 
digit to be passed through Vn. A coincidence of two or three 




Fig. 13-11. Three-input adder circuit using twin triodes. 



input pulses is sensed by the V^-Vs-Vq combination, which can 
cause a carry pulse to pass through V\ 2 and also inhibit the sum 
pulse by shorting Vn through F 9 and V\q. However, the latter 
pair can be prevented from shorting out Vn in case all three 
input leads receive input signals, whereupon tubes Vn and V% are 
both caused to conduct. In this case, Vi and V% cut off V$ and 
Fio in spite of their positive grids, and the original pulse to Vi, 
V 2 , and Vz is allowed to act through Vn. 

Another three-input coincidence-type adder may be derived 
from triode tubes connected in the parallel, or Rossi, coincidence 
circuit described in Chap. 4. As an example, in Fig. 13-11 a 
negative pulse on either input A, B, or C" cuts off Vu a and thereby 



280 



PHYSICAL COMPONENTS AND METHODS 



causes a sum pulse to appear at the anode of Vi 2 &. The three 
input leads are so connected to the six grids of F 4 , V&, and F 6 
that if any two of these receive a pulse, one of these tubes will be 
wholly (both sides) cut off. This causes at least one of the 
buffers V 9a , F 9 &, or Fio to conduct and thereby cut off Fi 2o . In 
this circuit Vn acts as a logical D and not (D and E); it will 
allow a sum pulse to pass if D alone receives a pulse, but will not 
let it pass if D and E both receive a pulse, as in the case where 
two or three input pulses are received at A, B, and C. When 



»-OUT 



C.P. 




CARRY 



Fig. 13-12. Three-input adder using triple-control-grid gate tube. 



three pulses are received, V7 and Vs a are cut off, thereby trans- 
mitting a sum pulse directly through Fg&. The carry pulse is 
passed through F 9 or F10 whenever two or three input pulses are 
received. 

A final example is given to illustrate the simplification which 
results from the use of a triple-control-grid gate £ube. As 
discussed in Sec. 14-3, this type of tube poses a difficult design 
problem at the present time. In Fig. 13-12 a 1 pulse received on 
A, B, or C Avill gate a clock pulse out of F5 as a sum. Two pulses, 
however, will be detected by V 7 , V&, or V$ and used to inhibit a 
sum pulse through Vi, V 2 , F 3 by holding their second grids 
negative. Two pulses will also place a carry pulse in a delay 
circuit. Three pulses are detected by Vi, which gates a sum 
pulse through F5 despite the inhibiting action of V-,, V%, and F9 
on Vi, V2, and V 3 . 

13-4-3. Algebraic Adders. So far, only adders of positive 
numbers have been considered. The algebraic adder adds or 
subtracts numbers according to their sign. An example of a 
method by which this can be accomplished is described and is 
illustrated in block diagram form in Fig. 13-13. 



ARITHMETIC ELEMENTS 



281 



First consider the rules of binary subtraction. It may be 
facilitated by handling a borrow as a negative carry to the sub- 
trahend. In other words, if the digit 1 in the subtrahend 
is to be subtracted from in the minuend, the difference is 1, 
and a borrow is created. This may be handled by adding 1 
to the next digit in the subtrahend. Thus a new subtrahend 
B' is created which may be handled as indicated in Table 13-4, 
where A — B = D and negative carry C. From this table the 

Table 13-4 



Original digits 


B +C" 


A - (B + C") 


Minuend 


Subtra- 
hend 


Previous 

negative 

carry 


New sub- 
trahend 


Inter- 
mediate 
negative 

carry 


Final 

negative 

carry 


Differ- 
ence 


A 


B 


C" 


B' 


C 


C 


D 





























1 


1 





1 


1 





1 





1 





1 


1 





1 


1 





1 


1 





1 

















1 


1 





1 


1 











1 


1 





1 











1 


1 


1 





1 


1 


1 



following rules can be formulated: 

D = (B f or A) and not (B f and A) 
C = (B' or C) and not (B' and A) 

In order to form the new subtrahend B' and the intermediate 
carry C ', a half adder may be employed to add B and C" . The 
difference D may also be obtained from a half adder since the 
rule pertaining to it states that D is the sum digit of the sum of 
B' and A. The final carry C is obtained by gating B' or C and 
not (A and B') through gate VI. A block diagram of the logical 
components which might be employed to obtain these conditions 
is shown in Fig. 13-13. 

Figure 13-14 illustrates a circuit arrangement to carry out 
the rules just stated. Tubes Vi, V 2 , V 3 , and F 4 , comprise a 
half adder which transmits a clock pulse representing B' to the 



282 PHYSICAL COMPONENTS AND METHODS 







I 








2 
ADDER 






1 








1 










In 




HI 










AND 




NOT 


















IZ 








I 


NOT 

II 


B • 


1 

2 

ADDER 


-B— * »~ 






m 








' — ►■ 
— »- 


OR 




AND 




C'V-— 


— C* 











Fig. 13-13. Subtractor, block diagram. 




Fig. 13-14. Subtractor, schematic diagram. 



ARITHMETIC ELEMENTS 



283 



second half adder, made up of Vs, V & , V h and F 8 . Vi also 
receives pulses from A and forms the logical A and B', which is 
used to suppress the A pulse through V 5 and the B' pulse through 
Vg when these two occur simultaneously. Vs is a clock-pulse 
gate for the difference digit D from F 5 or V& (but not both). 
Vb and Fe are a part of half adder I and also furnish the not (A 
and B') pulse in place of blocks II and III in Fig. 13-13. 

The analysis of the carry circuit is not quite the same in the 
schematic as in the block diagram, but it is equivalent. Vn 
receives B' on one grid and B' and A on the other. The polarity 
of the signals is such that the output represents B' and not (B f and 
A). Vq serves an or function by receiving this signal on one 
grid and C from the first half adder on the other. Thus the 
output of Vq is (C or B') and not (B f and A). The output of V$ 
gates a clock pulse through Fio, from which it emerges as the final 
carry C. 

Algebraic addition may be performed by either an addition or 
a subtraction operation, the sign of the sum depending on both 
the sign and the relative size of the numbers to be added. The 
rules of algebraic addition are shown in Table 13-5 for two 
numbers A and B. 

Table 13-5 



Relative sizes 
of A and B 


Sign of A 


Sign of B 


Operation 


Sign of 
sum S 


\A\ > \B\ 


+ 
+ 


+ 

+ 


Addition 
Subtraction 
Subtraction 
Addition 


+ 
+ 


\A\ < \B\ 


+ 

+ 


+ 
+ 


Addition 
Subtraction 
Subtraction 
Addition 


+ 
+ 



It may be seen that the operation to be performed is determined 
by the signs of two numbers. If the signs are alike, the operation 
is addition; if unlike, subtraction. Moreover, if the signs are 
unlike, the relative sizes of A and B determine which is to be 
subtracted from the other and the sign of the sum. 

If the sign of each number is represented by a binary digit 



284 



PHYSICAL COMPONENTS AND METHODS 



occupying the first space preceding the number and coincident 
in time with a clock pulse, a control circuit containing a half 
adder may be used on the signs of the two numbers to determine 
whether they are alike or different. It should be connected to 
detect (A or B) and not (A and B) so that a pulse is produced 
when the signs are different. This algebraic-sign-control circuit 
is represented by block I of Fig. 13-15 which illustrates the com- 
plete block diagram of an algebraic adder. By means of the sign- 



CONTROL — I 



rz 



SWITCHING 
CIRCUITS 



SUB 



© 



a 



m 



£}~ 



ADO 



TANK I 



B 



SUB 2 



TANK 2 



Q 



C.R 
Fig. 13-15. Algebraic adder, block diagram. 

control circuit a pair of switches in block II is set so that the 
digits are sent either to the adder III or to the two subtracters 
IV and V, depending upon whether the signs are the same or 
different. The three gates shown are used to remove the sign 
pulses accompanying some of the input digits. In the case of 
addition, it is necessary to remove only one of the two sign 
pulses and pass the other on to the sum because all signs are the 
same. 

To determine which number is the larger without using any 
extra time, tAvo subtracters are provided, and both subtractions 
are carried out simultaneously. The subtracter which is sub- 
tracting a large number from a smaller generates an extra carry 
pulse at the end of the arithmetic operation. This pulse is used 
to suppress the whole result before it is transmitted further. 



ARITHMETIC ELEMENTS 



285 



The sign of the minuend is introduced into each subtracter, but 
since the difference and the sign are suppressed in the circuit with 
the incorrect result, the proper sign is received at S along with the 
correct sum. 

13-4-4. Amplitude Adders. A group of adders can be com- 
pounded on the basis of the principle that discrete voltage levels 
may be created in various circuits by applying integral numbers 
of pulses simultaneously. The detecting circuits in the adders 




Fig. 13-16. Two-input amplitude adder. 

previously described respond to all voltages which are not below 
a critical value. The circuits described here, however, respond 
to pulses above a specified minimum but not above a critical 
maximum. Thus, a certain amount of simplification may be 
gained. This simplification, however, is realized only at the 
expense of greater effort to reduce circuit-component tolerances 
and maintain pulse amplitudes within prescribed limits. 

An amplitude adder compounded out of two amplitude half 
adders is illustrated in Fig. 13-16. Vi is biased so that it can 
be made to conduct by the application of one standard pulse, 
and V 2 is made to conduct by two standard pulses. Therefore 
Vi is the logical or and V 2 is the logical and for A and B. 7 3 is 



286 



PHYSICAL COMPONENTS AND METHODS 



conducting and biased so that it will be cut off by one or more 
pulses. Therefore it responds to A or B from V\. When both 
Vi and V 2 are caused to conduct, the two out-of-phase output 
pulses cancel at the grid of F 3 , and no pulse passes. A pulse 
through F 2 also sends a carry pulse through F 4 and into the one- 
pulse time-delay line: Vs and V& are identical to V\ and V 2 and 
comprise the second half adder, which adds the intermediate 
sum to the previous carry from the delay circuit. If V 2 does not 



out s 




Fig. 13-17. Three-input amplitude adder. 



put a pulse into Vi, there is a possibility that V& will pass a 
carry pulse to V A and thence into the delay circuit. V \ and V& 
never pass pulses simultaneously because if V 2 forms a carry, 
the intermediate-sum digit is zero, and F 6 does not receive a 
sum pulse from Va. This is because a sum and carry of 1 are 
mutually exclusive. 

Figure 13-17 illustrates a further simplification through the 
use of amplitude-sensitive elements. The circuit is the three- 
input type in which addend, augend, and carry are all received 
at the input. In this circuit, Vs responds to one or more positive 
pulses to cause a sum pulse to be gated through F 4 . V 2 responds 
to two or more pulses and prevents the sum pulse from leaving F 3 
by driving its second grid negative. At the same time, it causes 
a carry pulse to be gated through F 5 and into the delay circuit. 
Three simultaneous input pulses are passed by V\ as well as by 



ARITHMETIC ELEMENTS 



287 



the other two input tubes. V\ causes a sum pulse to be passed 
through Vi, and V 2 again initiates a carry pulse. 

Amplitude adders are of particular interest in ternary arith- 
metic (based on radix 3). Several schemes have been suggested 
for the physical basis of ternary arithmetic, and one having a 
nonclassical radix which is easy to deal with physically will be 
described. The three digits of a ternary system might be taken 
as +1, 0, and — 1, and these might be represented by a positive 
pulse, the absence of a pulse, and a negative pulse, respectively. 
The significance of the statement that the digits of a ternary 
system are +1,0, and — 1 is that the nth-order digit of a ternary 
number is +3 n-1 , 0, or — 3 n_1 , depending upon the digits +1, 
0, or —1 in the nth-order position. Such a representation is 
illustrated in Table 13-6. 

Table 13-6 



Decimal 


Normal ternary 


Nonclassical radix 





000 


o, 


0, 


1 


001 


o, 


0, +1 


2 


002 


o, 


+1, -1 


3 


010 


o, 


+ 1, o 


4 


011 


o, 


+1, +1 


5 


012 


+1, 


-1, -1 


6 


020 


+1, 


-1, 


7 


021 


+1, 


-1, +1 


8 


022 


+1, 


0, -1 


9 


100 


+1, 


0, 


10 


101 


+1, 


0, +1 


11 


102 


+1, 


+1, -1 


12 


110 


+1, 


+ 1, o 


13 


111 


+1, 


+1, +1 



If the positive and negative pulses are held at an equal and 
standard amplitude, several of the rules of ternary addition may 
be enforced rather easily. These rules are : 



Addend and 
augend 


Sum 


Carry 


+0 
+ (+l) 

o + (-i) 

( + 1) +(-D 




+ 1 
-1 











288 



PHYSICAL COMPONENTS AND METHODS 



To enforce these rules requires only that pulse currents be 
passed through a common load resistor and the voltage drop 
taken as the sum. This arrangement might be called a pulse- 
bucking adder. 

The remaining ternary addition rules are: 



Addend and 
augend 


Sum 


Carry 


( + 1) +( + 1) 
(-D + (-1) 


-1 

+ 1 


+ 1 
-1 



These two rules can be handled by amplitude circuits similar to 
those just described. Figure 13-18 illustrates a ternary, ampli- 
tude half adder which obeys the rules just stated. 



-vww 




Fig. 13-18. Ternary, amplitude half adder. 



ARITHMETIC ELEMENTS 289 

Vi and V 2 are normally biased to cut off and therefore respond 
only to positive pulses. Furthermore V 2 is biased to such an 
extent that it will begin to conduct only if a positive pulse is 
received at A and B coincidentally. When it does begin to 
conduct, it cuts off Vi at its second grid so that only gate Vi 
gets a pulse. Likewise in case a single positive pulse appears 
at the input, only gate tube F 3 receives a pulse. V z and F 4 
gate clock pulses into the S' and C output leads. As an example, 
when F 4 gates a clock pulse corresponding to ( + 1) + ( + 1), S' 
receives a negative pulse for the sum digit of — 1, and C receives 
a positive clock pulse for a carry digit of +1. 

In the lower half of the circuit, V$ and ^6 are normally con- 
ducting and respond only to single and double negative pulses, 
respectively. F7 and F 8 are pulse inverters, and V 7 receives an 
inhibitory pulse on its second grid to stop the pulse to clock-pulse 
gate V9 when two simultaneous negative pulses are received at A 
and B. F 8 is also a gate and passes a positive sum and a negative 
carry when two negative pulses are received at A and B. The 
problem of isolating the output circuits is not dealt with in this 
drawing. 

13-4-5. Coded-decimal Adders. It is possible to assemble 
gates and switches to perform addition in any number system 
which follows a known set of rules. Any coded-decimal system 
is an example of one such computational scheme. The 10 
decimal digits are given a code configuration, and arithmetic 
operations are carried out according to rules which account for 
the relationship between the code and decimal arithmetic. There 
are many possible code configurations, one group of which uses 
binary digits to represent decimal digits. (Thus there are four 
binary digits for each digit in the original decimal number.) 
One such code is known as the excess-3 code. In this code, 3 is 
added to each decimal digit to give an excess-3 value which is 
then represented by a corresponding four-place binary number. 
This code has the advantage that the 9's complement used in 
subtraction may be formed by transposing Ts and 0's. This 
feature of the excess-3 code is illustrated as an example of what 
can be done with various codes but does not indicate that it is 
the best code for all purposes. There are too many possible 
factors which are interrelated with machine details to be con- 
sidered in picking a code; therefore, to discuss abstractly the 
relative merits of various codes is not helpful. 



290 PHYSICAL COMPONENTS AND METHODS 

The following rules govern addition in the excess-3 coded- 
decimal system: 

1. A carry in the fourth binary place coincides with a decimal 
carry. As an example: 

Decimal notation: 

5 + 6 = 11 

Excess-3 binary notation: 

1,000 + 1,001 = 10,001 

2. When two coded-decimal digits in excess-3 code are added 
the sum is in an excess-6 code since each digit produced an 
extra 3. 

Thus two sets of digits can be added directly and checked to 
see whether a carry occurred in the fourth place or not. If it 
did, the sum may be returned to an excess-3 code by neglecting 
the fifth digit and adding 13. If no carry occurred, 3 should be 
subtracted from the sum to reduce it from excess-6 to excess-3 
code. 

The block diagram of a circuit which performs these checks and 
additions is shown in Fig. 13-19. It requires a minor-cycle pulse 
generator which generates a pulse in coincidence with every 
fourth clock pulse. 

Two excess-3 coded decimal numbers to be added are intro- 
duced serially into half adders I and II at A and B. The delayed 
carries produced by this addition are brought to gate Gi, which 
also receives minor-cycle pulses delayed one pulse time. The 
synchronization between the operand digits and the minor-cycle 
pulses is such that the latter reaches G\ at the same time as the 
fourth-binary-digit carry (if one occurs). Whenever a fourth- 
place carry takes place, flip-flop FF, which has just been reset 
by generator I, is set by means of a pulse through G\. Flip- 
flop FF in turn opens gate G 3 so that binary 1,101, or decimal 13, 
is introduced into half adder III in synchronism with the binary 
sum Sb of addend and augend. If no fourth-place carry occurs, 
FF opens G 2 , and binary 0,011 from generator II is introduced 
into half adder III. 

The second adder receives the sum from the first adder and a 
+3 or —3 from G 2 or G 3 , respectively. Since the latter pulses 



ARITHMETIC ELEMENTS 



291 



are delayed four pulse times after the former, the first sum is put 
through a four-pulse time-delay circuit before entering the second 
adder. The second adder performs in a normal manner except 
for a provision to inhibit a carry pulse from the fourth binary 
digit if it should occur. This is accomplished by passing all 
pulses through G 4 , which is open during all but the fifth pulse 
time when the delayed carry from the fourth place would be 




H ! H iiHi^rl 



1/2 

ADDER 
II 



■©-- 



MINOR 
CYCLE 
PULSES 

Generated: 

EVERY 

4 th PULSE 

TIME 




GEN. 



RESE" 



'/2 

ADDER 

III 



gi^^-Hgz 



FF 



GENERATOR 
— JGEN.II00II 



GENERATOR 



— |GEN. 1 101 
G.P. 



gH 



'/2 

ADDER 
IV 



NORM 
OPEN 



— fnp] — n pulse 

L — ' TIMES 



5P 



Fig. 13-19. Excess-3 coded decimal adder. 



received. Then the clock pulse closes the gate and eliminates 
this undesired carry. 

Figure 13-20 illustrates how the block diagram of Fig. 13-19 
might be effected with electronic tubes. Vi corresponds to 
gate Gi and V 2 is the flip-flop circuit which is set and reset by 
the two halves of V s . Vi is merely an inverter for reset clock 
pulses. The flip-flop controls gate tubes Vu and Vu correspond- 
ing to G 2 and G s . F 5 and F 7 comprise the +3 or 0,011 generator, 
while Vq, V s , and F 9 generate the 1,101 for —3. The sections 
of F10 are inverters which introduce the +3 and —3 into gate 
tubes Vu and Vu. Vu gates a clock pulse into the second adder 
for each pulse received from Vu or Vu. Vu and Vu form a gate, 



292 



PHYSICAL COMPONENTS AND METHODS 




ARITHMETIC ELEMENTS 293 

Gr 4 , and a pair of inverters, respectively, which inhibit the fourth 
carry pulse from the second adder as described before. 

13-4-6. Other Types of Adders. Addition may also be per- 
formed on a diode matrix similar to the code conversion tables 
described in Chap. 3. No significant simplification seems to 
result from the use of such adders. 

Some work has been done to develop electronic adder tubes 
which operate with cathode-ray beams and anodes arranged to 
produce sum and carry pulses for proper beam deflections. 
Although such tubes may ultimately result in a more economical 
and compact unit than arrangements using conventional tubes, 
the latter are sufficiently flexible and economical to ensure their 
use for some time to come. An adder tube was developed at 
RCA during the Second World War. 6 

13-5. Elementary Ring -type Adders 

A ring-type adder is essentially a counter which makes sums 
by counting out the digits of two numbers to be added. Decimal 
ring-type adders are used in the ENIAC and are described in 
Sec. 4-5. For binary addition, the counter ring may be reduced 
to a single flip-flop like the one illustrated in Fig. 3-2 on page 
15. It is first cleared, and then the corresponding digits of 
addend and augend are presented to its input in turn. The sum 
is produced modulo 2, and another flip-flop or its equivalent is 
employed to store the carry. 

Since the ring-type adders must receive one operand before 
the other, they are not well suited for use in serial adders of the 
type described in the previous section. They are, however, a 
popular form of adder for parallel arithmetic units. 

13-5-1. Parallel Ring Accumulator. An accumulator is a coun- 
ter which adds to its contents each number transmitted to it. 
An accumulator made of binary rings will count both numbers 
and their complements. Figure 13-21 represents an accumula- 
tor of this type. When it is desired to transfer a number from 
the register to the accumulator, all the G\ gates are opened, and 
the transfer is made, leaving the original number stored in the 
register. If it is desired to transfer the complement of the num- 
ber in the register to the accumulator, all the Gi gates are opened. 

At least two methods are available for detecting and handling 
carries in a parallel adder with ring-type elements. In one 



294 



PHYSICAL COMPONENTS AND METHODS 



method the carry is stored in a flip-flop which has a gate asso- 
ciated with it. The output of each gate is connected to the 
input of the binary counter to the left. The carry is begun by 
pulsing all carry gates. Since each carry may initiate another 
carry, the process must be repeated until the carrying process 
is finished. If the adder contains n counters, n pulse times 
normally have to be allowed in order to assure that all carries 
are effected. As explained in Sec. 13-3, the average number of 



ACCUMULATOR F F 



ADD 
PULSE 



SUBTRACT, 
PULSE 



REGISTER 





F 


F 




T" 









pti 






.rf . 




H 


c 


> 1 





Fig. 13-21. Binary accumulator. 

carries is well under n, so, on the average, less time is required 
for an automatically sequenced computer which can continue 
as soon as all carries are complete. A circuit connected to each 
flip-flop stage in such a way that it detects the presence of zeros 
in all the flip-flop stages can be used to sense the completion of 
the carry. 

A second method of adding allows all carries to be made simul- 
taneously, thereby enhancing the characteristics of nonautomatic 
sequence computers. An adder of this type is described in the 
next section. 

13-6. Larger Arithmetic Units 

It is usually desirable to code operations, other than addition 
and subtraction, so that they can be handled automatically. 
These other operations, however, are compounded out of the 
basic addition operation. Several complete adders and multi- 
pliers are described in this section in order to indicate the types 
of arithmetic units which might be devised. 

13-6-1. Serial Arithmetic Units. A serial adder using any 
one of the coincidence adder units described in Sec. 13-4 is illus- 



ARITHMETIC ELEMENTS 



295 



trated in Fig. 13-22. If a three-input adder is used, provision 
must be made to return carry pulses to one of the input leads 
at the next pulse time. The serial adder employs two storage 
units into which electrical pulses are entered serially, and from 
which they emerge at some later convenient time. Mediums 
suitable for this purpose, such as transmission lines or rotating 
drums, are discussed in Sec. 14-2. The box R contains the 
amplifiers and clock-pulse gates required to regenerate these 
pulses after each trip through the storage medium. The com- 
bination of storage unit and regenerating unit forms a device for 



c 



ADDEND A ND SUM 

STORAGE A 

Gl 



OUT IN "G3W^ 



STORAGE B 
AUGEND 



IN ^G3 

At)D SIGNAL 



D 



- Tr> 



AJ 



ADDER -^ 



G2 



-®- INVERTER 



IN 

Fig. 13-22. Serial adder. 

retaining digital data; such a device is defined as a register. 
The shaded circle designated I, upon receiving a signal to begin 
the addition operation, provides switching potentials to gates 
Gi, G 2 , and (7 3 . Obviously, the delay time in the storage section 
of a serial register must be equal to or greater than the time 
required for all the serial digits in a binary number. 

If addend and augend are first entered into their storage units 
A and B, respectively, and an add signal is applied as the lowest 
order digits are passing into R, consecutive higher order digits 
from A and B will be summed in the adder. G 3 is opened by the 
add signal, so addend digits do not continue to circulate through 
A but are replaced by sum digits, which enter as addend digits 
leave. Thus addition with a serial adder requires one cycle 
time and leaves one operand circulating in one register and the 
sum in the other. 

With three registers having circulating times twice that 
required for introducing the operands, a serial binary multiplier 



296 



PHYSICAL COMPONENTS AND METHODS 



may be developed from the serial adder described. A device 
of this sort is illustrated in the block diagram of Fig. 13-23. 
This multiplier performs additions of the multiplicand, the 
additions being controlled by successive digits of the multiplier. 
The partial sums are circulated in storage A. The multiplier 
and the multiplicand are entered into their respective C and B 
registers. As the two lowest order digits emerge from the storage 
registers, a clock pulse is applied to reset flip-flop FF and to 




NO OF DIGITS 
IN OPERANDS 
P = PULSE TIME 



STORAGE C 



n 

OUT IN 



C.P EVERY 
PULSE TIME 



Fig. 13-23. Serial multiplier. 

open gate G 5 . If the multiplier digit is a 1, it sets FF, which in 
turn opens 6r 4 , which adds all the digits of the multiplicand into 
register A . After 2n + 1 pulse times, the second-order digit of the 
multiplier is at G 5 , and the first-order digit of the partial product 
in register A is entering the adder. If the second-order digit of 
the multiplier is a 1, the multiplicand is again added to the con- 
tents of the product register. This time, however, the mul- 
tiplicand is delayed by one pulse time so that a shift of the 
number by one digit order is effected before addition with the 
preceding partial product. In this manner the process is 
repeated, the multiplicand shifting one place for each circulation 
and being added to the partial product each time a 1 appears in 
consecutive higher order positions of the multiplier. 









ARITHMETIC ELEMENTS 297 

Although the serial multiplier requires a minimum of equip- 
ment, the total multiplication time is the product of the circula- 
tion time times the number of digits in the operands. As an 
example, it would take 1,800 pulse times to multiply two 30-digit 
binary numbers in the manner described. Some saving in time 
might be obtained by rounding off the digits in the product, but 
one circulation would still be required for each digit in the 
multiplier. 

13-6-2. Parallel Arithmetic Units. The basis of the parallel 
arithmetic unit is either the elementary ring- type adder of Sec. 
13-5 or the coincidence adder of Sec. 13-4. The former type has 
been studied by the Raytheon Company, 10 and the results of this 
investigation are used as the basis for a description of a parallel 
adder with simultaneous carry. A modification of the Raytheon 
adder which would allow multiplication is outlined briefly in 
order to indicate the degree of complexity of a parallel multiplier. 

Parallel Adder with Simultaneous Carry. An adder in which 
all carries are accomplished simultaneously has been proposed 
by the Raytheon Company 10 and is illustrated in Fig. 13-24. 
The gates labeled G\ are employed to introduce the addend digits 
into the accumulator. The presence of a carry is indicated by a 
1 in the addend register and a in the accumulator and is detected 
by C?4, which in turn controls G 2 . This latter gate receives a 
carry pulse which, if the gate is open, is distributed to the delay 
line following each stage of the register. The carry pulses which 
pass G 2 are also inverted in I and sent to G 3 , which senses the 
coincidence of a in the accumulator stage and an open (x 4 gate. 
This indicates a double carry and means that a carry pulse should 
be passed on to the next stage. If 6r 3 is open, that is exactly 
what happens, so all carries are made with one carry pulse. 
Actually, n rise times are required to make n carries, and in the 
Raytheon circuit, about five pulse times have to be allowed on 
this account. This is a great deal faster than the serial adder, 
which requires about n 2 pulse times for addition of two n-digit 
numbers. 

Shifting in a Parallel Ring-type Register. In order to multiply 
numbers in the conventional manner of summing the partial 
products, it is necessary to shift the accumulated partial product 
with respect to the new partial product before the two are 
summed. A shifting register consisting of one flip-flop for each 



298 



PHYSICAL COMPONENTS AND METHODS 




ARITHMETIC ELEMENTS 



299 



digit stored is illustrated in Fig. 13-25. One plate of each flip- 
flop is connected through a delay circuit to the grid of the 
opposite tube in the next flip-flop. A bus delivers reset pulses 
to the remaining grids simultaneously. When a reset pulse is 
applied to these grids, the flip-flops which are storing a 1 are 
returned to the condition, and a pulse is introduced into each 
delay line following a flip-flop in which a 1 was stored. At the 
end of the delay period, the pulses emerge from the delay lines 
to set the following flip-flops. Thus, binary numbers stored in 

END AROUND CARRY BUS 



DELAY CIRCUITS 

i 

REGISTER 1 |FF| 

r 


P 

1 


h 

P 

t 


P 

1 


h 

t 


p p ip 
r 3? 

t 


IN 
RESET 


IN 


IN 


IN 




PULSE 
















j D 


1 — -DELAY CIRCUIT 



Fig. 13-25. Shifting register. 

this register are shifted one space to the right each time a reset 
pulse is applied. 

Another way of shifting is to provide two registers and pro- 
vision for shifting numbers from one to the other. By providing 
a set of gates for each type of transfer, the digits may be shifting 
to the right, to the left, or not at all as the number is passed from 
one register to the other. 

In some operations it is desirable to retain digits which would 
be shifted out at one end of the register during the shifting 
process by entering them into the other end of the register as they 
leave. As an example, in the parallel shift register of Fig. 13-25, 
the digits shifted out of FF n could be returned to FFi through a 
delay circuit and the dotted line shown. 

Parallel Multiplier. A multiplier might be devised using the 
parallel adder and the shift register which are described above. 
A three-digit multiplier is illustrated in Fig. 13-26. The mul- 
tiplicand is placed in the upper flip-flop register, and the mul- 
tiplier is placed in a serial register. The product accumulator 
is twice the length of the operand registers and is also a shift 
register. 



300 



PHYSICAL COMPONENTS AND METHODS 




ARITHMETIC ELEMENTS 301 

When FF 4 is set by 1 pulses from the multiplier register, it 
opens gates (75 and G Q for one addition time (about five pulse 
times). During this time an add and a carry pulse are applied 
to adder gates G\ and G 2 so that the multiplicand is added into 
the number in the accumulator. The timing of the various steps 
is such that the lowest order digit of the multiplier is introduced 
into FF 4 first. If it is a 1, the multiplicand is added into the 
accumulator. After the addition time, the partial sum in the 
accumulator is shifted to the right, FF4 is reset, and the second 
multiplier digit is applied to FF±. The process is continued in 
this manner until all the digits of the multiplier have been used. 
At that time, the number in the accumulator is the product of 
the two numbers multiplied. In the example shown, the last 
digit of the product is also cut off. 

It may be seen that the time required to multiply two w-digit 
numbers is n addition times. 

REFERENCES 

1. Alt, F. L., A Bell Telephone Laboratories Computing Machine — I, 
Mathematical Tables and Other Aids to Computation, Vol. 3, No. 21, pp. 
1-13 (January, 1948). 

2. Burks, A. W., Electronic Computing Circuits of the ENIAC, Proceedings 
of the Institute of Radio Engineers, Vol. 35, pp. 756-767 (August, 1947). 

3. Burks, A. W., H. H. Goldstine, and J. von Neumann, Preliminary Dis- 
cussion of the Logical Design of an Electronic Computing Instrument. 
Report prepared under Contract W-33-034-ORD-7481 between the 
Research and Development Service, Ordnance Department, U.S. Army, 
and the Institute for Advanced Study, Princeton, N.J. (June 28, 1946). 

4. Hilbert, D., and W. Ackerman, Grundzuge der Theoretische Logik (J. 
Springer, Berlin, 1928). 

5. Mauchly, J. W., Preparation of Problems for ED VAC Type Machine, 
Annals of the Computation Laboratory of Harvard University, Vol. XVI, 
pp. 203-207 (1948). 

6. Rajchman, J. A., R.C.A. Computron, Final Report, LC PB 40611 (Office 
of Technical Services, Department of Commerce, March, 1943). 129 pp. 

7. Tar ski, Andrew, Introduction to Logic (Oxford University Press, New 
York, 1939). 

8. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, Report on the ENIAC, Technical Report I (University of Penn- 
sylvania, Philadelphia, June 1, 1946). 2 vols. 

9. University of Pennsylvania, Staff of Moore School of Electrical Engi- 
neering, Progress Report on the EDVAC (University of Pennsylvania 
Philadelphia, June 30, 1946). 2 vols. 

10. West, C. F., and J. E. DeTurk, Digital Computer for Scientific Applica- 
tions, Proceedings of the Institute of Radio Engineers, Vol. 36, No. 12, pp. 
1452-1460 (December, 1948). 



CHAPTER 14 
TRANSFER MEDIUMS 

14-1. Introduction 

Computing consists of performing arithmetic operations on 
numbers. In the strictest sense, numbers are abstractions; 
arithmetic operations are performed on representations of these 
abstractions or numbers. The numbers themselves are defined 
by the rules for performing the arithmetic operations. For 
example, the mark 8 in the decimal system and the mark 100 in 
the binary system are representations of the same number. Any 
computing machine operates, in accordance with applicable rules, 
upon representations such as these. 

In the following discussion number is used in a broader sense. 
It is used to mean an ordered set of digits considered as a unit, 
regardless of the meaning of the digits. For example, some of a 
particular set of digits (the first 20, for example) may be the 
binary representation of a particular number. Some others in the 
set may represent an address code and still others in the same 
set an instruction to the machine. We shall call the entire set 
a number. Thus, we are using number as a synonym for what 
others in the computing-machine field have termed a word. 

In this chapter we will discuss the means for transferring 
numbers in space and time during the course of a computation. 
Any connected series of arithmetic operations implies the neces- 
sity for such transfers, and the concept of transfer is basic in any 
discussion of computing machines. To facilitate discussion and 
to provide a basis for grouping transfer systems for comparison 
purposes, the definitions set forth below have been adopted. 

Transfer in time alone is necessary when a number is to be 
reused later at the same location in space. Transfer in both 
time and space is necessary if the number is to be reused later 
at a different location in space. As noted below, some types of 
time transfer are properly spoken of as storage. Therefore, it 
may be seen that the terms transfer and storage, as used with 

302 



TRANSFER MEDIUMS 303 

respect to computing machines, are related. Transfer, the 
broader of the two, includes storage, because transfer includes 
any change in location with respect to either time or space, or 
both. The term storage is used with respect to a transfer in time 
only. Moreover, not all transfers in time are properly referred to 
as storage operations. A number is in storage when its representa- 
tion is introduced into a medium with the intention that it shall 
remain there until specifically called for and withdrawn. The 
number is merely being delayed (as contrasted with being stored) 
when it is introduced to some medium from which it is auto- 
matically delivered after a fixed interval of time. This dis- 
tinction is illustrated in the following examples: 

Suppose it is desired to transform the serial representation of a 
15-digit number to a parallel representation. This transforma- 
tion may be visualized by imagining 15 digits proceeding in 
single file down one channel and then turning simultaneously 
at right angles, to proceed down 15 separate parallel channels. 
In practice, however, the transformation from serial to parallel 
expression requires the delaying of each digit by an appropriate 
interval. The first digit is delayed for the longest time, the 
second for the next longest, etc. When the delays are properly 
adjusted, the last digit reaches the entrance to its particular 
channel just as each of the other 14 digits is ready to be shunted 
into its particular channel. 

Each of the first 14 digits in this case has been transferred in 
time, but the delaying operation is not thought of as storage. 
Storage is the term used, rather, when data are made to exist in 
some interpretable form until called for. An elementary storage 
is accomplished by means of a pencil and a piece of paper. A 
legible mark on a piece of paper remains in interpretable form for 
use whenever needed, so long as the paper is not destroyed and 
the mark is not erased. The problem in calculating with the 
aid of pencil and paper is not to delay information but simply 
to store it for later reuse. In a computing-machine operation 
it may be necessary to store information in interpretable form for 
later reference (perhaps for many later references), and it may be 
necessary also to delay particular data for the purpose of carrying 
out a particular computation. The example of the serial-to- 
parallel conversion illustrates the latter. 

Transfer systems used for storage or delay purposes are referred 



304 PHYSICAL COMPONENTS AND METHODS 

to as either static or dynamic. In a dynamic system, the data 
are in motion, usually cyclically, relative to the medium in which 
they are represented. An example of the dynamic system is the 
sonic delay line ; in this system, trains of sound waves, represent- 
ing digits, travel through a substance in which the velocity of 
sound is slow as compared with the velocity of electrons in a 
conductor. The waves are introduced at one end of the delay 
line and received at the other. The same signal may recirculate 
indefinitely through the delay line (being retransmitted at one 
end each time it is detected at the other), but it is always in 
motion relative to the reading and writing elements, i.e., the 
receiving and transmitting transducers. 

In a static system the stored data remain fixed with respect to 
the storage medium until transferred by the operation of some 
switching device. A system composed of electromagnetic relays 
is an example of a static system. A rotating magnetic drum also 
is a static storage system, according to this definition, because 
data recorded on the drum are static with respect to the surface 
of the drum. However, the data on the drum are in motion 
relative to the magnetic heads which read and write upon the 
surface; hence, in one very important respect, the static mag- 
netic-drum system and the dynamic delay-line system are 
equivalent. If we define as a cyclic storage system any system 
in which each stored datum continually moves in cycles with 
respect to the reading and writing stations, then both the rotating 
magnetic drum and the delay line are cyclic storage systems. 
Using the converse as the definition of a noncyclic system, a bank 
of relays is an example of this type. 

If we define the access time of a system as the time required 
to withdraw a number from storage, the essential difference 
between cyclic and noncyclic systems is the following: In a cyclic 
system, the average access time is closely related to the time inter- 
val between successive appearances of the same storage cell at a 
location where its contents can be withdrawn. In a noncyclic 
system, the speed of operation of the switching device is impor- 
tant. The relationships between these factors and other 
design considerations are discussed below. Another difference 
between cyclic and noncyclic systems which is worth noting 
as an aid in narrowing the definition is the following: In a non- 
cyclic system, the transfer of data in space from the storage 



TRANSFER MEDIUMS 305 

medium to some other part of the computer involves noncyclic 
motion in space, i.e., motion of relay contacts, of an electron 
beam, or of some other physical mechanism. In a cyclic system, 
this transfer of data to or from storage involves noncyclic motion 
in time while a cyclic motion in space is taking place. This 
motion in time is the delay which takes place between the 
instant an order related to a particular storage cell is activated 
and the instant that cell, in the course of its cyclic motion, 
arrives at the proper reading or writing station. 

Another pair of mutually exclusive terms, one of which may be 
applied to any storage system, is the pair erasable and nonerasable. 
These terms are essentially self-explanatory. For example, data 
stored as combinations of punched holes in a tape or a card are in 
nonerasable form; data represented by the condition of mag- 
netization of a magnetic surface are in erasable form. 

There is one more pair of terms which will be found useful in 
discussing storage systems: volatile and nonvolatile (or permanent). 
These terms refer to the degree of physical stability of a repre- 
sentation of stored information. In a volatile storage medium, 
like a delay line, retransmission of each signal once during each 
storage cycle period is required. In case of power failure, the 
stored data are lost. If the data are recorded on magnetic 
material, they are in permanent storage, because the representa- 
tion of each digit, a particular magnetic condition of a per- 
manent magnet material, will not deteriorate even if the com- 
puting machine is turned off; an external force is required to 
eradicate the data. 

14-1-1. Cyclic Storage Systems. Sonic delay lines and mag- 
netic drums are probably the best known of the cyclic storage 
systems. Other types have been used successfully, e.g., elec- 
tromagnetic delay lines. These are discussed in detail in other 
sections of this chapter. The use of radar transmissions to 
distant objects and the deflection or intensity modulation of 
electron or ion beams also have been suggested. In theory, at 
least, any device for sending a signal over a path relatively 
long as compared with the signal used to represent a digit is 
applicable in a cyclic, dynamic, storage system. 

The basic physical characteristics of any cyclic storage system, 
with respect to each separate single-digit channel, or path, in the 
system, are: 



306 PHYSICAL COMPONENTS AND METHODS 

1. The storage cycle, which is 

a. The physical length of the dynamic storage path, 
divided by 

b. The speed of a digit in storage, relative to the receiving 
and transmitting (or " reading" and "writing") stations. 

2. The number of digits which can be carried on the storage 
path at one time, which is 

a. The physical length of the dynamic storage path, divided 

by 

b. The physical length of the signal representing a single 
digit in the storage medium. 

3. The radix, which is the number of elementary states dis- 
tinguishable in the medium. 

Primarily upon these characteristics depends the quantity of 
information which can be stored and the speed with which the 
computing mechanism can have access to each stored datum. 

For purposes of illustration, let us consider a magnetic-drum 
storage system. A binary digit is represented on the surface of 
the drum by the magnetic state of a particular area. If this 
area is magnetized in one direction, the digit it represents is 1. 
If the area is magnetized in the other direction, the digit is 0. 

In practice, 1,600 inches per second has been found to be a 
reasonable drum surface velocity (corresponding to a storage 
concentration of 80 digits per inch). A small drum, capable of 
storing 1,024 digits in each single-digit path, or track, has a 
storage cycle of about 8 milliseconds. (The physical length of 
the storage path is actually a little more than 1,024 times 0.0125 
inch — which is the physical dimension of the digit, measured in 
the direction of motion — because a small amount of dead space 
must be left on the drum.) In practice, a drum 4.3 inches in 
diameter will carry 1,024 binary digits per track. 

Let us apply the same considerations to a mercury storage 
system. The velocity of sound in mercury is about 5.7 X 10 4 
inches per second (1.45 X 10 5 centimeters per second 48 ). For a 
storage path 14 inches long, the storage cycle time will be 

14 

^-= X 10 -4 = 250 microseconds 
5.7 

Suppose the digits are introduced to storage at the rate of four 
digits per microsecond, which is the rate at which the BINAC, 



TRANSFER MEDIUMS 307 

built by the Eckert-Mauchly Computer Corporation, operates. 
The physical length of the signal representing each digit in 
storage is therefore 

5.7 X 10 4 X 0.25 X lO" 6 S 0.014 inch 

The number of digits carried on the storage path at one time is 

14 



0.014 



1,000 



Mercury delay lines are suitable for the storage of a few 
thousand digits. The problem of controlling the temperature 
(in order to maintain constant velocity) is not difficult to solve 
in a delay line of reasonable length, e.g., of the order of a thousand 
digits. 

It is not practical to store a very large number of digits in a 
mercury system because of the number of vacuum tubes required. 
For purposes of estimating engineering complexity, it may be said 
that approximately 11 tubes are required to generate signals and 
control and read each delay line in the EDVAC. At the rate of 
11 tubes per thousand-digit line, 11,000 tubes would be required 
for a million digits. In terms of cost and complexity this would 
be a high price. On the other hand, such a large storage system, 
with the access speed of a thousand-digit delay line, is not likely 
to be required. (The engineering aspects of the mercury storage 
design problem are discussed in more detail in Sec. 14-2-6.) 

In a cyclic storage system, the average time required to locate 
and read from or write into any given storage cell may be related 
directly to the storage cycle period. This average access time 
can be reduced by adding more reading and writing stations. 
For example, several magnetic recording heads can be placed at 
equal intervals around a magnetic drum to reduce the average 
access time without reducing the storage capacity or speeding 
up the drum. However, the price paid for this reduction in 
average access time is high, in terms of increased equipment 
complexity. Applying this same technique to a mercury delay 
line is not practical because the additional stations introduce 
serious attenuation. 

Where greater storage is required (e.g., storage for 1 million 
digits) the mercury delay line is not practical, without serious 
modification to increase its capacity by eliminating the effect 



308 PHYSICAL COMPONENTS AND METHODS 

of temperature variation on sound velocity, because of the 
number of vacuum tubes required. At the rate of 11 tubes per 
1,000 digits, 11,000 tubes are required for 1 million digits. On 
the other hand, mercury delay lines are well suited for storing a 
few thousand binary digits. 

The following seem to offer some promise of immediate and 
future applicability in cyclic storage systems in computing 
machines: 

1. Sonic delay lines, liquid medium. 

2. Sonic delay lines, solid medium. 

3. Magnetic recording. 

4. Electromagnetic delay lines. 

5. Punched tape, read electrically. 

6. Punched tape, read photoelectrically. (Note possibility 
of punching holes with electric arc.) 

7. Photographic tape, read photoelectrically. 

8. Phosphor drums. 

9. Electrostatic storage tubes, read cyclically. 

14-1-2. Noncyclic Storage Systems. The important con- 
siderations pertaining to noncyclic storage systems are: 

1. Radix — the number of distinguishable states which the 
medium may assume. 

2. The complication of the switching equipment — electro- 
magnetic, electron beam, or other. 

3. The complication of equipment required to set the storage 
medium in its various radix states and of the equipment 
required to detect these states (write and read). 

The types of noncyclic static storage which now seem to be of 
interest are: 

1. Relay — electromagnetic or electron tube. 

2. Electrostatic storage — cathode-ray tube or Selectron. 

As indicated above, in meeting operational requirements with 
noncyclic storage systems the main difficulty lies in providing 
switching facilities. If the choice of items of intelligence stored 
or to be stored is by means other than physical motion of the 
storage medium, and if items corresponding to even a few thou- 



TRANSFER MEDIUMS 309 

sand binary digits are to be stored, it is clear that the problem of 
switching between the storage cells may become awkward. 

A convenient, well-known, rapid means of switching utilizes 
the deflection of beams of electrons or the blocking of beams of 
electrons by electrostatic fields. This means gives promise of 
becoming useful in connection with computers; this is particu- 
larly true since other ingenious techniques have been developed 
to utilize this switching when the number of leads is small 
enough to compare favorably with the number required for 
dynamic storage. This system is described more fully in Sees. 
14-2-8 and 14-2-9 on electrostatic storage schemes. 

In cyclic storage, as opposed to noncyclic storage, the switching 
function may take place mainly on a time basis. This immedi- 
ately raises a host of problems in connection with synchronization, 
and it implies some sort of periodicity with consequent limitations 
on the speed with which the system can function. The same 
objections apply to systems in which switching is on a space 
displacement basis. Maximum velocities attainable imply a 
minimum consulting period. Thus, unless the information is to 
occur and to be used in an orderly and predictable manner, any 
cyclic storage system is wasteful of time. Noncyclic systems are 
intrinsically much better suited for consulting where the addresses 
of the information stored are likely to be somewhat random. 

It would be possible to use as noncyclic storage most of the 
cyclic storage schemes which do not involve transmission of 
intelligence as wave motion or some similar motion through a 
static medium. Punched tape, for instance, could be used in a 
static noncyclic storage application, but this would involve a 
tremendous increase in the number of reading stations as well 
as complications in switching. 

Early noncyclic, static storage systems frequently used an 
array of switches. Both mechanical relay systems, with some 
scheme for stabilizing and holding each relay in position, and 
vacuum-tube relays arranged in flip-flop circuits were used. 
In either case, storage of binary digits is simple and straightfor- 
ward. There are no difficulties other than switching difficulties. 

In exactly the same way, punched tape, photographic tape, 
and other similar systems could be used instead of relays. Again 
the only difficulty would be in switching. Magnetic storage 
also could be used in a noncyclic application (though radical 



310 PHYSICAL COMPONENTS AND METHODS 

modifications in the system of reading would be required) because 
the state of magnetism in an object is easily read by means of a 
coil moving relative to the magnetic medium. 

The use of magnetic material seems almost automatically to 
imply the idea of motion, primarily in cyclic systems. However, 
three different schemes are available for reading the state of 
magnetism of a body without moving the body itself: (1) a 
coil on a magnetic core, or a magnetic head, may be moved past 
the magnetized object and thus furnish the same relative motion 
which would be used to detect the magnetic condition of the 
body if the body itself were moved; (2) the nonlinearity of the 
magnetization curve may be used, so that an increase in mag- 
netomotive force in a particular direction changes the flux sub- 
stantially only if the substance is not saturated in that direction ; 
(3) a cathode-ray beam may be deflected by the magnetic field of 
the substance. Of these schemes only the second seems to show 
any promise at all, and no well-developed efforts to utilize this 
scheme are known to have been carried out. However, much 
work has been done recently on satur able-core reactors, and 
use of some types of these in magnetic storage is feasible. 

14-2. Transfer in Time 

Examples of the need for transfer of numbers in time are 
given in Sees. 5-3 and 14-1. Several time storage mediums are 
discussed in the following material, and for convenience of com- 
parison they are considered in application to: 

1. The high-speed storage section of a computer. 

2. Permanent storage. 

3. Supplementary storage. 

In a given computer there may be some overlapping of these 
types of storage, but each has its own distinctive features. 

The high-speed storage section of a computer includes the 
registers of the arithmetic unit, and its requirements are inti- 
mately tied in with the type of arithmetic chosen. The most 
important criteria for high-speed storage mediums are (1) their 
access time; (2) the flexibility with which they can be switched 
from one unit to another; and (3) their simplicity and reliability. 

Permanent storage is desirable if the computer is to operate 
with the aid of tables, repetitive program routines, repetitious 



TRANSFER MEDIUMS 311 

similar input data, and output data from one problem as input 
data for another problem. The choice of permanent storage 
medium determines certain characteristics of the input-output 
equipment and usually is chosen with that point in view. Input- 
output systems are hereinafter discussed as an important type of 
permanent storage system. The principal criteria for judging 
a permanent storage medium are (1) the compactness of the 
medium, or the number of digits stored per unit cost and per 
unit volume; (2) the life expectancy of the secondary medium 
on the shelf and in operation; (3) the rates at which data may be 
written in and read from the medium ; (4) the ease of handling the 
medium mechanically by the machine, and by human hand when 
being removed from or loaded into the machine; (5) the ease 
with which the data on the medium can be changed for additions, 
deletions, or corrections; (6) the facility with which the data 
can be reproduced as a new permanent storage medium; (7) the 
suitability for convenient inspection and checking; and (8) the 
reliability of the medium and its associated equipment. 

There is usually a need for supplementary storage having 
characteristics which are a combination of those of high-speed 
storage and of permanent storage. This is the result of the 
expense in complexity of extending high-speed storage to cover 
all the computing requirements and the relatively poor accessi- 
bility of numbers in permanent storage when the numbers 
required must be read in a random order. Criteria for good 
supplementary storage mediums are (1) the flexibility and speed 
of switching for reading from and writing on the medium; 
(2) the capacity of the medium in number of digits stored as a 
function of volume, cost, and complexity; (3) erasability and 
reusability of the medium; (4) reliability of the medium and its 
associated equipment. Since access time and capacity of a 
supplementary storage system are closely related, (1) and (2) are 
necessarily considered together. It is imperative that supple- 
mentary storage systems be erasable and reusable. 

14-2-1. Relay Storage. The only forms of high-speed relay 
storage employed in present-day large-scale digital computers are 
the electromechanical 1,2 or the electron-tube relay. In the 
Harvard Mark II and the Bell Telephone Laboratories machines, 
electromechanical relays are employed, whereas in the ENIAC, 
electron-tube relays, or flip-flop circuits, are employed. The 



312 PHYSICAL COMPONENTS AND METHODS 

latter type is also employed in the EDVAC and other electronic 
computers. Although they are relatively large and expensive, 
relays are also used in the first three computers as supplementary 
storage mediums. The present tendency is not to use relays as 
supplementary storage for large amounts of data. They are 
well suited, however, as the high-speed storage medium in 
conjunction with arithmetic units. 

Admissible Radix. The operation of the electromechanical 
relay as a switch is described in Chap. 4. In present-day com- 
puters, two stable positions of this type of relay are utilized so 
that these relays store digits modulo 2. It is possible, however, 
to establish three stable states in a relay equipped with centering 
springs and a pair of coils which deflect the armature in opposite 
directions. As a variation of this configuration, a permanent 
magnet might be made for the armature so that its deflection is 
dependent upon the direction of current flow through a single 
coil. Thus it is possible to construct an electromechanical relay 
which could be used to store radix 3 numbers. The conventional 
electron-tube flip-flop circuit has but two stable states. 

Access Time. The pickup, or setting, time for commercial 
electromechanical relays used in computers varies from 1 to 
about 10 milliseconds. The relays used in the Mark II equip- 
ment operate in from 6 to 10 milliseconds. The wire-brush 
type of relay used by the IBM Corporation contains up to 
12 contacts and operates in 2 milliseconds. Another relay dis- 
cussed in Chap. 3 has been made to operate in 1 millisecond. 
The total access time for digits in the Mark II is 16% milli- 
seconds. This is longer than the pickup time of the relays and 
includes factors of safety for the synchronization between units 
and current-switching equipment described later in this section. 
It seems likely that 10 milliseconds is about the minimum access 
time which can be expected when electromechanical relays are 
used. 

The time duration of a digit in a flip-flop circuit may be very 
short but is limited by the figure of merit of the tubes. Flip-flop 
circuits cycle satisfactorily in 0.1 microsecond and may be 
expected to have a reliable access time of 0.2 microsecond. 

Mode of Operation. A desirable feature associated with the 
electromechanical relay is the simplicity of the auxiliary equip- 
ment required to generate digital signals. A relay may be set 



TRANSFER MEDIUMS 313 

by a d-c pulse, the wave shape of which is not critical. Reading 
information out of a relay requires merely the detection of the 
position of the contacts and may be done instantaneously. The 
armature may be held either by the current flowing through the 
coil or by a mechanical latch which is released by a second coil. 
In the latter case, the storage unit is said to be nonvolatile because 
it can sustain loss of power without losing all the information 
stored. 

If it is desirable to have the contacts of the relay carry large 
amounts of current, a synchronous system may be employed in 
which the relay contacts are opened and closed before current is 
applied to them. This is done in the Harvard Mark II machine, 
in which pulses of direct current are generated by large cam- 
operated switches and distributed to the relays as signal pulses. 
Thus, the only current-making and -breaking contacts are at 
one point where they are available for inspection and service. 
This form of operation, however, has the disadvantage that some 
time is wasted between the relay operation time and signal- 
current time. 

The electromechanical relay is versatile in its application as 
an arithmetic component because it has a pair of low-impedance 
input or actuating leads which may be kept isolated completely 
from those of the output. 

Reliability. Reliability in computing devices is of the utmost 
importance, and therefore a serious problem arises with the use 
of electromechanical relays. Interference with one or more 
computing operations may result if a particle of dust happens to 
rest on the contact faces, since electrical contact is made between 
two points of microscopic size. Since there is no direct way to 
detect such a failure either before or after it occurs, the use of 
checking circuits is desirable. Fortunately, it is relatively easy 
to add contacts to an electromechanical relay for checking- 
purposes. This was done on the Bell Telephone Laboratories 
machines, and their ability to resume operation quickly after a 
failure is one of their outstanding features. 

A comparison between the electromagnetic relay and the elec- 
tron tube reveals that the former has a longer life expectancy in 
operating hours but will operate fewer times than the latter. 
Furthermore, the electromechanical relay is subject to failure 
from vibration and dust on the contact faces and is apt to require 



314 PHYSICAL COMPONENTS AND METHODS 

more delicate care than a vacuum tube. The relay usually 
operates on less power than the filaments of a vacuum tube, 
but the latter can operate on alternating current taken directly 
from the power mains through step-down transformers. Most 
fast-acting relays operate on direct current; hence when the 
rectification conversion loss is taken into account, their power 
requirements are comparable to those of electron tubes. The 
two are also nearly the same in cost and size. The greatest 
difference in characteristics is in the maximum speed of opera- 
tion ; the electron-tube relay is at least 10 4 times faster than the 
electromechanical relay. Therefore, in situations where com- 
puting may be done serially, a single tube can perform the same 
operations in the same time as a large number of relays operating 
simultaneously in parallel. 

Size and Cost. The cost of relays of the type used in the 
high-speed storage section of a computer varies between $1 and 
$5 per digit stored, depending upon the amount of auxiliary 
equipment which is used. The required volume per digit stored 
is about 3 to 6 cubic inches. 

In small computing devices, relays are particularly suitable 
because of their availability and versatility. On the other hand, 
they have also been used quite lavishly in large computing 
machines such as the Harvard Mark II computer, which employs 
a total of about 13,000 relays. These relays are made by Auto- 
call, and most of them are of rugged construction. They have 
six-pole, double-throw contact arrangements with large silvered 
contacts, consume about 6 watts at 100 volts direct current, and 
operate in 6 to 10 milliseconds. About 5,000 of these relays 
of the latch type, Model HHA-JD, are used in the internal stor- 
age. There are about 9,000 relays in the largest Bell Telephone 
Laboratories machines. Most commercially available electro- 
magnetic relays are described in a book compiled at MIT, 1 and 
flip-flop circuits are amply covered in the references for Chap. 3. 

14-2-2. Punched-tape Storage. A common storage medium 
for digital information is found in tape which is opaque to light 
or is physically strong enough to operate mechanical feelers 
and has a low electrical conductance. Information is stored by 
means of an array of holes which are placed in the tape and 
located in reference to sprocket perforations. 3-7 Tape is used for 
storage in the Bell Telephone Laboratories and the Harvard 



TRANSFER MEDIUMS 315 

machines as well as in many special-purpose devices. Because 
it is nonerasable, it is most suitable as a medium for permanent 
storage. It may be used in rolls or, where iteration is desired, 
in endless loops. 

Admissible Radix and Arrangement. An area on a tape of the 
sort discussed above may have two possible elementary states, 
punched and not punched. The admissible radix is therefore 2. 
Several adjacent columns, or data channels, may be punched 
along a length of tape. 

Access Time. The limiting factor which determines the 
minimum spacing between holes may be the tendency of the tape 
to shrink or stretch, the precision with which the holes can be 
punched in the tape, with respect to some reference such as a 
row of sprocket perforations, or the minimum spacing at which 
holes can be reliably sensed by the reading equipment. The 
minimum time interval between digits is directly proportional 
to the minimum space between holes and inversely proportional 
to the maximum linear velocity at which the tape can safely be 
moved with accurate guidance. 

Two types of reading equipment are in present use: electrical 
sensing and photoelectric sensing. Electrical sensing of holes in 
punched tape is accomplished by passing the tape over an elec- 
trically charged metallic platen, over which are placed metal 
wipers in position to make contact with the platen through holes 
in the tape, or by arranging fingers which slip through the holes 
in the tape and operate switches. The top speed to be expected 
with electrical sensing is about 100 digits per column per second. 
The output is in the form of electrical pulses, which are con- 
venient for computer use. 

The top speed at which data can be recorded on standard tele- 
type tape by the Western Union 10B Reperforator is 10 operations 
per second. Since as many as five holes may be punched at 
once, this corresponds to a speed of 10 X 5 or 50 digits per 
second. The Model 12A transmitter will read 75 digits per 
second. 

A faster method of sensing holes in punched tape is by means 
of photoelectric cells. The tape is driven past a mask which 
contains optical apertures of proper size and spacing for one hole 
in each row. These apertures are illuminated through the holes 
in the punched tape by a constant source of light. Light which 



316 PHYSICAL COMPONENTS AND METHODS 

passes through the holes and the apertures is gathered in an 
optical system which is arranged to focus the light on photocells. 
An electrical pulse is generated in a photocell each time a punched 
hole passes over an aperture in the mask. With this type of 
system, the British have succeeded in sensing holes in punched 
tape at a rate of 5,000 digits per second per column. A five-hole 
tape was used, so this corresponds to a sensing rate of 25,000 
binary digits per second. 

Mode of Operation. It is sometimes convenient to use tape as 
a cyclical storage medium by forming it into a continuous loop. 
In the Harvard Mark II equipment these loops are used to store 
tables of functions and other data which are to be used repeatedly. 
Punched tape is also used as a supplementary storage medium 
in the Bell Telephone Laboratories machine. By punching 
data into the tape at one station and reading them again at 
another, a short-time storage unit is made available. This fea- 
ture proved to be much more useful than originally anticipated. 
A practical limit to the number of digits which may be stored by 
this method is the mechanical difficulty of handling a long loop 
of tape. A cyclical series of digits can be created in this way, 
however, with no complication in reading and recirculating 
equipment, because punched tape is provided with sprocket 
holes which automatically serve to maintain synchronism 
between the tape and the read-write equipment. In order to 
shorten the access time of the system, it is quite feasible to use 
independent read and write heads at various locations around 
the loop, thereby adding only the complication of additional 
read-write circuits and switches. Presently available tape- 
preparing equipment employs mechanical punches actuated by 
electromagnets and operated from a keyboard similar to that 
of a typewriter. The electromagnet circuits have been modified 
to operate from the outputs of computers but are inherently 
limited in speed by their physical inertia. The possibility of 
perforating tape with an electric arc is of interest because of the 
higher perforating speeds which might be attained. 

If it is desired to correct or change data punched on tape, 
it is necessary to cut the tape, add or remove sections, and splice 
it together again. This process is tedious but is simplified by 
the fact that the recorded data can be inspected visually and 
directly. This factor results in a simplification of the checking 
process when errors are found to exist in the storage medium. 



TRANSFER MEDIUMS 317 

Reliability. An important advantage of punched tape at the 
present time is the fact that tape-preparing and -reading equip- 
ment has been in commercial use for many years and has been 
refined to a high state of reliability. The wearing character- 
istics of this tape depend upon the method of reading, the amount 
of acceleration required in driving it, and the type of drive 
mechanism. Photoelectric sensing can be effected without use 
of fingers or pins pressing against the tape, thereby eliminating 
one source of wear. In any case, a paper tape has a fairly short 
life owing to wear on the sprocket holes even when it is driven 
at constant velocity. On the other hand, standard commercial 
equipment is available with which to reperforate duplicates of 
the master tape for spares so that they can be replaced as they 
become worn. 

Teletype tape is reliable to the extent that it can operate under 
a reasonable range of temperatures and is unaffected by small 
amounts of dust. Since sprocket or reference holes are punched 
in the tape as the data are added, the synchronization problem 
is minimized. The Bell Telephone Laboratories machines use a 
six-hole tape and a code in which two of the six holes are used 
for each digit. As a check against additional holes which might 
appear in the tape as a result of wear, a circuit is provided which 
stops the machine in case more or fewer than two holes are 
detected simultaneously. 

Size and Cost. There are so many materials from which 
punch tape can be made and so many varieties are now available 
that it is beyond the scope of this book to discuss each sepa- 
rately. However, the common five-hole teletype transmitting 
tape is described as a representative sample, and some data 
concerning other tapes are listed in tabular form at the end of 
Chap. 17. The capacity of teletype tape is about 3 X 10 7 binary 
digits or punches per cubic foot, and the cost of the tape is 
trivial: about one dollar per 3 X 10 6 binary digits. 

Punched tape is commercially used in several different forms. 
The Wheatstone tape used for Morse code representation is 
about \"2 inch wide and accommodates two rows of dots and a 
row of sprocket perforations. This tape costs about 20 cents per 
roll and may be procured through the Link Paper Co. in New 
York. 

Teletype tapes are available in five-, six-, and seven-hole 
widths and may be procured at about 25 cents per roll through 



318 PHYSICAL COMPONENTS AND METHODS 

manufacturers, the names of whom will be supplied by the Tele- 
type Corporation upon request. The five-hole tape which is 
common to office equipment may be procured at stationery stores 
at about 75 cents per 1,000-foot roll. A tape of greater durability 
than ordinary paper tape is identified as gray fiber tape and has 
been procured through Western Union by special arrangement. 

There are also opaque tapes, such as film leaders, available 
from the Eastman Kodak Company and other film manufacturers, 
which might be used for punched-tape application. Thirty- 
five-millimeter film leader costs about 2 cents per foot when 
purchased in quantity. 

Commercial Tape-punching and -reading Equipment. The 
Teletype Corporation makes printers, tape readers, and tape 
punches of several styles. This equipment is highly refined and 
has been in reliable use for many years. The following equip- 
ment is incorporated into the Harvard Mark II computer: 

1. Teletype Model 15, Page Printers. 

2. Teletype Model 10B, Reperforators. 

3. Teletype Model 12A, Transmitters (five-hole). 

4. Teletype Model 11 A, Transmitters (six-hole). 

5. Teletype Model 101, Motor-Driven Reperforators (both 
five- and six-hole). 

6. Teletype Model 3B, Distributor-Transmitter (five-hole). 

7. Teletype Model 10A, Distributor-Transmitter (six-hole). 

8. Distributors, stepping switches (or rotary switches), relays, 
and associated gear. 

The operating speeds of some of these pieces of equipment are 
of interest. Ordinarily, the printers operate at between 390 and 
460 operations per minute, although they will operate as fast as 
600 operations per minute. The transmitters normally operate 
at 900 operations per minute on a five-wire system. This equip- 
ment is used by Western Electric and Western Union. 

Equipment for perforating and reading the two-hole Wheat- 
stone tape used in automatic telegraphy is available through the 
McElroy, Kleinschmidt, and Creedy Companies. The McElroy 
mechanical tape-reading equipment operates at speeds up to 
500 words per minute. Two pairs of binary spaces are used to 
represent a Morse code letter, so there are approximately 25 pairs 
of binary spaces per word. 



TRANSFER MEDIUMS 319 

The IBM Corporation also uses tape-perforating and -reading 
equipment which is available on a special rental basis. Much of 
this equipment is adapted to transferring data from punched 
cards to tape and back again. 

Drive Mechanisms. In the Bell Telephone Laboratories 
machines a variable-speed tape drive is used, and tables are 
indexed by sections. Therefore, when searching for an entry in a 
large table, the section may be designated, and the tape drive 
drives the tape at high speed until the desired section is reached, 
whereupon the speed is decreased to normal reading speed so that 
the entry may be withdrawn. This feature complicates the drive 
mechanism considerably, so that the cost of punched-tape 
equipment depends to a considerable extent on the drive mecha- 
nism employed. 

14-2-3. Photographic Storage. Storage of digital data on 
photographic strip film is similar to storage on punched tape in 
many respects and may, therefore, be classified with it in the 
group of nonerasable storage mediums. 8-12 Data may be repre- 
sented on photographic film by areas of different opacity which 
are applied photographically to a film base and which are detected 
photoelectrically. Work is being carried out on the utilization 
of film as a nonerasable storage system by O'Neal and Tyler of 
the Eastman Kodak Company. This has been described in the 
literature. 10-12 Most of the information to be presented on this 
subject has come from these studies. 

Recording of digital data may be effected by illuminating the 
film through a mask which restricts the images of a bank of 
discharge tubes or the screens of cathode-ray oscilloscope tubes. 
The discharge or c.r.o. tubes act as light sources which may be 
modulated by the signals to be recorded. In accordance with 
the coded information supplied to the modulators of the light 
sources, beams of light are varied in intensity to produce a hori- 
zontal row of clear and opaque blocks. After one row has been 
exposed, the film can be advanced and another row of data 
composed and photographed. The film may be moved con- 
tinuously if the flashing time of the light source is made short 
compared to the time required for the film to travel the aperture 
distance of the mask. 

Admissible Radix. Although it is possible to record and 
distinguish between several degrees of opacity on film, present 



320 PHYSICAL COMPONENTS AND METHODS 

efforts are being concentrated on a two-state representation of 
data in which the light is either passed or cut off. This on-off 
technique has been considered necessary for perfect dependa- 
bility of operation. 

Access Time. The digit area or block area must be con- 
siderably larger than the limit set by the resolving abilities 
of the film emulsion. A more important operational criterion is 
the requirement for optical reliability when the film is slightly 
dusty or somewhat scratched. Coded areas 0.010 by 0.020 inch 
are of the magnitude considered feasible. It is probable that 
dots of this size may be recorded at a rate of 10 4 elements per 
second per channel. If 50 channels are placed side by side on a 
35-millimeter film, this rate would result in a total recording 
speed of 5 X 10 5 digits per second. 

Reading is accomplished in a manner similar to that described 
for the photoelectric reading of punched tape. A bank of photo- 
cells is located on one side of a slit which is illuminated from the 
opposite side through the film. An optical system is arranged to 
project light from each block section of the slit to a corresponding 
photocell. Thus, opaque areas of the film interrupt the light to 
corresponding photocells, and transparent areas allow light to 
pass through. Electric currents generated in the photocells 
from varying illumination are then amplified to form the output 
signals. Photoelectric reading equipment is limited in speed 
only by the rise time of the photocell and its amplifier and has an 
inherent speed of at least 10 6 centimeters per second; furthermore 
it is capable of reading independently of scanning velocity. 
The maximum reading speed achievable is limited by the accuracy 
with which the film can be guided past the reading stations and 
is in the order of 10 4 elements per second per channel, not 10 6 
centimeters per second. 

Mode of Operation. The minimum requirements for film 
recording include a light source which can be modulated at a 
high repetition rate, an optical system capable of fine resolution, 
and a film-advancing mechanism capable of guiding the film 
accurately. Flash tubes of the type used in commercial stro- 
botacs are being investigated for their applicability to this field 
as light sources. 8 Cathode-ray tubes with low-persistence 
screens are also being studied. 11 A single c.r.o. tube is used to 
generate a row of marks across the film by stepping the electron 



TRANSFER MEDIUMS 321 

stream across the face of the tube, the beam being modulated so 
that spots of light appear at the desired locations behind a mask 
on the tube face. 

A straightforward method of recording a binary number and its 
inverse simultaneously, using a c.r.o. tube, has been demon- 
strated. If a mask with two horizontal rows of apertures is 
placed over the face of the c.r.o. tube, the vertical sweep may be 
used to choose one of each pair of spots in a given column, and 
the horizontal deflection may be used to scan the spot of light 
over the columns in sequence. Thus, when the light spot is 



PATH OF BEAM 
NUMBER -^ 



□ 



\ 
\ 
\ 



COMPLEMENT 



□ >-»"- □ 






NUMBER IOt 
Fig. 14-1. Binary optical recorder. 

recorded in the number, a dark spot is recorded in the cor- 
responding column of the inverse row, and vice versa. Several 
mask patterns are possible, and one is illustrated in Fig. 14-1. 

Film, like punched tape, may be used in rolls and in endless 
loops. Since it is essentially a permanent storage medium, it is 
most useful for storage of tables and as an output medium. 

Reliability. Dependable driving equipment for standard- 
width strip film has been developed for the motion-picture 
industry. 

The life expectancy of the film depends upon the method by 
which it is driven. Ordinary film has a reasonably long life 
when driven with sprocket wheels. The life expectancy is 
increased if the film is driven with rubber rollers instead. When 
the sprocket holes are not used, however, a synchronizing channel 
must be recorded on the film. As the film ages, copies may be 
made by standard commercial copying procedures. Minor 
changes may be made by the usual film-cutting and -splicing 
techniques. Film may be inspected visually with the aid of 
film-viewing apparatus. 



322 PHYSICAL COMPONENTS AND METHODS 

The reliability of this medium depends largely on the reliability 
of the handling equipment and on the degree to which the film 
can be kept clean and unscratched. A positive signal for each 
digit and a check on reading equipment may be had by recording 
the inverse of each number below the number. This allows a 
positive signal for both O's and Ts and provides the basis for a 
good checking system, for in the film reader each digit of the 
number and its inverse can be compared. Failure of the record- 
ing equipment to record a digit or failure of one reading section 
is shown in the comparison process. 

Size and Cost. Data may be stored on photographic film 
to a density of 10 9 binary digits per cubic foot and at a cost of 
10 6 binary digits per dollar. Thirty-five-millimeter positive 
film is available at about $3.75 per 200-foot roll, and commercial 
processing averages about $3 more per 200-foot roll, depending 
upon the quantity. Films with faster emulsion speeds cost 
two or three times as much as positive film. 

A disadvantage of film as a storage medium is the complexity 
of photographic processing. Automatic equipment is available 
Avhich speeds up the conventional developing procedure and 
which continuously turns out strip film at a rate of 6 feet per 
minute. An even faster process may result from work now 
being done by the Haloid Corporation in connection with 
Xerography, an electrostatic and thermal dye-deposition process. 

It is not known that any equipment of the type described is 
now commercially available, and its future cost is difficult to 
estimate. One of the largest factors in cost is the type of film 
drive required. In some cases a servodrive is desired which will 
search for and locate values in a table. Driving the film at a 
variable velocity introduces the problems of film-roll acceleration, 
servo stability, and the provision of a reference sensing equipment 
with a link to the drive mechanism. 

The use of film as the medium on which the final output of a 
computation may be recorded is of interest and is described in 
Chap. 15. Film can be exposed at a high speed and can record 
charts and graphs as well as numbers and coded data. Enlarged 
prints can be made on paper for visual inspection. 

14-2-4. Magnetic Storage. Magnetic recording 17 is a tech- 
nique which presents attractive possibilities for both the internal 
storage and the input-output functions of computing devices. 






TRANSFER MEDIUMS 323 

The alterable physical state requisite to any storage medium 
is in this case the intensity of residual magnetization (magnetic 
moment per unit volume) in a small volume of magnetic mate- 
rial. The techniques used in the magnetic recording of speech 
and music have been treated extensively in the literature. 19-39 
These sources should be consulted for background material and 
fundamental principles. The material in this section will 
attempt to indicate factors peculiar to the magnetic recording 
of digital information. 

The Magnetic Recording Process. The magnetic recording 
process briefly is as follows: To record information, the medium, 
or carrier, is caused to move past the gap of a recording head, 
which is a suitably designed electromagnet energized by the 
signal current. Each succeeding element of the carrier is brought 
into a definite magnetized state, in which it tends to remain upon 
leaving the recording field. The recorded pattern of remanent 
magnetization along the carrier is related, in a manner deter- 
mined by the system characteristics, to the time variation of the 
signal current. This pattern is read by drawing the carrier 
past the gap of a reproducing head, which may be similar to the 
recording head. Some of the magnetic flux from the section of 
the carrier in close proximity to the gap follows the low-reluctance 
path through the core of the head. Time variation of this flux 
generates an output voltage in the coil surrounding the core. 

Reading can be accomplished by means of saturable-core 
reactors, also, as indicated on page 310. 

In the recording and reading of digital information, this output 
voltage need not be a replica of the recording signal voltage. 
It is sufficient that the system recognize in a reliable manner the 
value of the digit stored at a given location on the medium. 
For practical reasons, this type of recording is generally limited 
to binary digits, where the system need recognize only two 
possible values of digit. 

The magnetizable medium, or carrier, may take a variety of 
physical forms. It may be a homogeneous metallic wire or tape. 
The magnetic material may be an alloy of appropriate magnetic 
properties, plated on the surface of a ductile, nonmagnetic wire 
or tape. One increasingly popular form of carrier is the coated 
nonmetallic tape. A suspension of a powdered magnetic iron 
oxide (magnetite) is applied in a thin layer to a paper-tape base. 



324 PHYSICAL COMPONENTS AND METHODS 

In another form the powder is dispersed uniformly through a 
plastic tape. In internal storage applications it is generally 
desirable to record on the surface of a rigid member such as a 
drum or cylinder. In such cases the rigid member may be made 
of a magnetic alloy, the magnetic material may be plated or 
coated on the surface, or a magnetic tape may be bonded to the 
surface. 

The choice of physical form of the carrier depends mainly on 
the requirements imposed by the intended application. At one 
extreme, for example, there is the simple input application, in 
which information is recorded at a relatively low speed and at a 
later time read out at a higher speed in synchronism with an 
internal storage unit. There is no accessibility-speed or searching 
requirement, as information is read out only in recorded sequence. 
The record must be capable of permanent storage. A large 
quantity of information must be stored on a single record. The 
erasability feature is not essential. For this type of application, 
the medium is best handled in the flexible form of fine wire or 
thin tape wound on a pair of spools or reels. 

The rapid internal-storage type of application lies at the other 
extreme. Here the information may have to be recorded, read, 
and erased at frequent intervals. It must be possible to locate 
quickly any specified storage cell for transferring information to 
or from the cell. A reasonable arrangement for this type of 
service is a rapidly rotating rigid drum with the magnetic medium 
on the cylindrical surface. 

Information may be recorded on a number of parallel tracks 
located side by side along the length of the drum. In this way, 
the time required for a selected cell to reach the reading head 
will not exceed one revolution period of the drum. If several 
reading heads are spaced around the circumference of each track, 
this maximum waiting time is correspondingly reduced. 

Between these two extremes many variations in physical form 
are possible. For example, a supplementary storage system of 
moderate reference speed might use as a medium an endless 
loop of tape containing two or more parallel tracks of information. 

The styles of head most commonly employed in magnetic 
recording are diagrammed in Fig. 14-2. In Fig. 14-2a the pole 
pieces are in line with each other and are in contact with either 
side of the carrier. This magnetizes the medium in a direction 



TRANSFER MEDIUMS 



325 



normal to its surface and is known as perpendicular recording. 
In Fig. 14-26, the pole pieces are not in line but are displaced 
slightly with respect to each other. This arrangement tends to 
concentrate or focus the flux in the tape into a narrow region, 
so that wider pole tips of lower reluctance may be used. The 
resultant magnetization in the medium is still essentially per- 
pendicular, although it may have a small longitudinal com- 
ponent. In Fig. 14-2c, the pole pieces are still further displaced 
to provide longitudinal recording, in which the magnetization is 



(o) 



(b) 



(c) 






(a) 


(b) 


(O 


(d) 


MAGNETIZATION 


PERPENDICULAR 


PERPENDICULAR 


LONGITUDINAL 


LONGITUDINAL 


FORM OF CARRIER 


TAPE 


TAPE 


TAPE, WIRE 


TAPE, WIRE, DISC.DRUM 


CLEARANCE 


CONTACT 


CONTACT 


CONTACT 


CONTACT, NON-CONTACT 



Fig. 14-2. Magnetic-recording-head styles. 

parallel to the direction of motion. Figure 14-2<i shows a ring- 
type head in contact with one side of the carrier, also for longi- 
tudinal recording. This is an increasingly popular form of 
head. 

It is evident that round wire is satisfactory only for longitudinal 
recording. It is not suitable for perpendicular recording because 
there is no way of returning the wire to its original axial orienta- 
tion for playback. Tape is satisfactory for either method. 

Efficiency of magnetic coupling between the carrier and the 
head, and also sharpness of resolution, are more easily achieved if 
the pole pieces are maintained in contact with the carrier. Con- 
tact recording, however, is feasible only if the linear speed of the 
carrier is relatively low, for abrasion at the head due to high-speed 
passage of the carrier is sufficient to wear it away rapidly. This 
method is used almost exclusively in the recording of sound, 
where speeds fall in the range of 6 or 8 inches per second to several 
feet per second. Digital recording, on the other hand, generally 
calls for considerably higher linear speeds. Particularly for 



326 PHYSICAL COMPONENTS AND METHODS 

rapid internal storage systems, speeds upward of 100 inches per 
second must be used, and 2,000 inches per second is conceivable. 
At such high speeds, abrasion of the head and the carrier makes 
contact recording quite impractical, and it becomes necessary to 
maintain a small head-to-carrier clearance. Head styles a, b, c, 
and d in Fig. 14-2 may all be operated in contact with the carrier, 
but of these only the ring type of longitudinally magnetizing 
head d is considered suitable for noncontact recording. 

For rapid internal storage, then, we may confine our discussion 
to the case of noncontact ring-shaped heads and longitudinally 
recorded tracks around the periphery of a cylindrical drum. 

Transfer of Data. Transfer of information to and from the 
drum consists in general of three processes: erase, record, read. 

The purposes of erasure are to obliterate previously recorded 
material and to bring the medium into the initial magnetic state 
required by the recording process. The required initial state 
may be either complete demagnetization or longitudinal satura- 
tion. (A possible alternative to demagnetization is transverse 
saturation.) 

Erasure by saturation, generally referred to as d-c erase, is 
accomplished by subjecting the moving carrier to a strong con- 
stant magnetic field in the longitudinal direction. The d-c erase 
head may be either a d-c electromagnet or a permanent magnet 
of suitable design. 

Erasure by demagnetization, or a-c erase, is accomplished by 
the familiar method of subjecting each element of the carrier to 
an alternating magnetic field, the amplitude of which diminishes 
to zero over a number of cycles. This may be done with an 
a-c erasing head with a field which tapers off suitably on the 
trailing edge. If the demagnetization of a given element must 
be completed within a short distance of travel, then the erasing 
field must alternate at a high frequency. This may limit the 
depth of penetration of the erasing field so that previous signals 
are not entirely removed. Thickness and electrical resistivity 
of the magnetic coating of the carrier must be chosen to take 
this effect into account. 

The nature of the recording and reproducing, or reading, 
processes may be illustrated by first considering a very simple 
type of recorded signal and the signal forms associated with it. 
Suppose the carrier has previously been demagnetized by a-c 



TRANSFER MEDIUMS 



327 



erasure ; then if the carrier is held stationary with respect to the 
recording head while a signal is imprinted by passing a direct 
current briefly through the head winding, the resulting typical 
distribution of magnetization J along the carrier will be as shown 
in Fig. 14-3a. 

If the carrier bearing such a signal is moved with constant 
linear speed past the gap of the reading head, the shape of the 
curve of time variation of the flux threading the winding will be 



(o) 





(b) 



Fig. 14-3. Magnetization J plotted against linear displacement along the 
carrier. In (a) is shown a distribution produced by applying a d-c signal to 
the winding while the carrier is held stationary. In (b) is shown the effect 
which would be produced by applying a signal while the carrier is in motion. 



roughly similar to one of the humps shown in Fig. 14-4a. The 
electromotive force developed in the winding is equal to the 
product of the time derivative of this flux times the number of 
turns, and it will therefore have a shape like that shown in Fig. 
14-4fr. This voltage may be further differentiated by means of a 
simple resistance-capacitance network to give a single sharp 
pulse with definite polarity of the sort shown in Fig. 14-4c. 

Suppose now that the signal is recorded dynamically rather 
than statically, i.e., with the medium in motion. If the mag- 
netizing field were able to rise abruptly and remain constant for 
a brief interval T, then drop sharply to zero, the pattern of 
Fig. 14-3a might be expected to " stretch," as in Fig. 14-36. 
The flux pulse would display a flat top of length equal to the 
pulse duration times the carrier velocity. The second derivative 



328 



PHYSICAL COMPONENTS AND METHODS 



of this signal would contain a pair of pulses, which for most 
purposes would be undesirable. 

In practice, the magnetizing field will rise and fall in a manner 
determined by the input signal and the circuit and head constants. 
If the field is caused to fall from a fairly steep portion of the rise 
curve, the shape of Fig. 14-3a can be approximated to a reasonable 
degree in the dynamic case. Indeed, if a system is capable of 
yielding the flat-topped voltage pulse of Fig. 14-36, then the 



AA 



W 



» rv= : 



A 



X 



V 



(a) 



(b) 



(c) 



1 10 I VAC 

Fig. 14-4. Effects produced by motion of carrier on which a representation 
of the digits 110100 has been recorded. The signal shown is called a two- 
level return signal because the flux <j> varies between two levels (one of which 
represents the digit 1 and the other 0) and because it returns to the 0-level 
between 1-pulses. 

information represented by the pulse could evidently have been 
recorded in a shorter length of track. 

There are several ways in which binary digits can be repre- 
sented by a recorded signal. Three classes of signal are illus- 
trated in Figs. 14-4, 14-5, and 14-6. These represent the two- 
level return signal, three-level return signal, and the two-level 
nonreturn signal, which are discussed below. The horizontal 
coordinate in all of these represents time. The distance between 
adjacent vertical lines represents the time interval, or cell, 
allotted to one binary digit. A series of seven cells is shown. 
These contain the digit sequence 1, 1, 0, 1, 0, 0, with no signal 
entered in the seventh cell. 






TRANSFER MEDIUMS 329 

Figure 14-4a shows the wave form of the flux in the core of the 
reading head, for the case where the flux remains at a given level 
M to indicate either a or no signal, but shifts momentarily to 
another level N to indicate a 1. To a first approximation this 
time plot of (f> resembles the space plot of the magnetization 
intensity / along the carrier. The two levels M and N between 
which the flux shifts may correspond to saturation of the carrier 
in the positive and negative directions, respectively. That is, 
the carrier is previously d-c erased to saturation in one direction, 
and a 1 is recorded by pulsing to saturation in the opposite 
direction. Alternatively, the carrier may be a-c erased, in which 
case the level M corresponds to demagnetized carrier and zero 
flux, with JV corresponding again to saturation. (Although this 
would appear to provide about half the signal voltage, there is 
the possibility that the noise in a suitably demagnetized medium 
may be sufficiently lower than the noise for the saturated state 
to result in a net gain in signal-to-noise ratio.) 

The emf induced in the winding of the reading head is shown 
in Fig. 14-46. If this voltage is differentiated by passing it 
through a suitable resistance-capacitance coupling network, the 
derivative voltage will be of the form shown in Fig. 14-4c. The 
portion of this voltage which falls below the threshold indicated 
by the dotted line is a sharp pulse, the presence or absence of 
which denotes whether the digit being read is a 1 or a 0. 

Since only two states, or levels of magnetization, of the medium 
are employed, one of the two kinds of binary digit must neces- 
sarily be represented in the same manner as the absence of a 
digit. In Fig. 14-4a, is indicated in the same manner as no 
signal. In some applications this may be considered a serious 
limitation. 

In the three-level return system, a definite signal is used to indi- 
cate either kind of digit, with absence of signal reserved to denote 
a vacant cell. When three-level recording is used, the carrier 
is initially demagnetized by a-c erasure. A 1 is recorded by 
saturating the carrier in one direction with a pulse, and a is 
recorded by saturating it in the opposite direction. Definite 
positive and negative pulses are thus available for indicating l's 
and 0's to the output circuits. This method is illustrated in 
Fig. 14-5. 

In the two cases described above, the reading-head flux returns 



330 



PHYSICAL COMPONENTS AND METHODS 



to the no-signal condition at the end of each unit interval. 
Figure 14-6 illustrates a method of recording known as the 
two-level nonreturn system, in which the flux shifts from one level 
to the other only when the new digit is different from the one 
in the preceding cell. 

Figure 14-6a is the time plot of the recording-head flux <f>, 
which, again, is similar to the space plot of the magnetization J 
along the carrier. The recording head is driven by a flip-flop 
circuit in such a manner that the magnetizing field saturates the 



N + 

M 

L+ 





VW^W- 



(a) 



(b) 



(c) 



I I I VAC 

Fig. 14-5. Three-level return signal. 



carrier in the positive direction for one condition of the flip-flop 
and in the negative direction for the other. A 1 pulse at the 
beginning of a cell causes the flip-flop to shift to the positive 
condition, unless it is already there; a pulse, to the negative 
condition, unless it is already there. 

If the time required for the transition, as seen at the reading 
head, is taken as the minimum practical length of unit interval, 
then 4> will vary with time somewhat as shown in Fig. 14-6a. 
The voltage e\ induced in the reading head will vary as the time 
derivative of <£, as shown in Fig. 14-66. A positive pulse denotes 
the beginning of a series of l's and a negative pulse the beginning 
of a series of 0's. 



TRANSFER MEDIUMS 



331 



When ei swings beyond either the positive or the negative 
threshold level indicated by the dotted lines, it causes a reading 
flip-flop to shift to the corresponding positive or negative condi- 
tion, where it remains until stimulated by a pulse of opposite 
sign. The voltage e 2 appearing at one of the plates of the flip-flop 
pair is shown in Fig. 14-6c. 



(a) 



A 




\ 

j 


/ 

A 

f 

1 

(I 


\ 

V 

1 
J 


1 


1 









(b) 



(c) 



(d) 



(e) 






I I I VAC 

Fig. 14-6. Two-level nonreturn signal. 

Figure 14-6d shows a clock pulse e 3 of short duration occurring 
near the end of each interval. If e 2 and e 3 are applied to two 
grids of a gating tube, the clock pulses will be transmitted only 
when e 3 is in the positive state. The gate output voltage e 4 is 
shown in Fig. 14-6e, where the presence of a pulse denotes a 1 
and the absence denotes a 0. 

Comparison of Figs. 14-4 and 14-6 reveals some advantages 
of nonreturn recording. If the cell repetition rate is taken to 
be the same in both cases, it is evident that the maximum fre- 
quency which must be handled by the recording and reading 
heads in the nonreturn case is of the order of half the correspond- 



332 PHYSICAL COMPONENTS AND METHODS 

ing maximum frequency in the return case. Furthermore, if the 
length of carrier per cell is the same in both cases, the shortest 
region of unidirectional magnetization is about twice as long as 
the shortest region in the return case. This implies that if the 
heads and carrier in a given system are capable of storing a 
definite maximum number of digits per inch of track and of trans- 
ferring at a rate of so many digits per second on a return-signal 
basis, then the same system should store effectively twice as 
many digits per inch and transfer twice as many digits per second 
on a nonreturn basis. 

Transfer Schemes. The specific manner in which digits are 
transferred to and from tracks on a storage drum is subject to a 
degree of choice. 

Consider the type of recording designated as selective mark 
insertion, in which the two magnetic states of the carrier are 
saturation in the positive and negative directions, on either a 
return or nonreturn basis (Figs. 14-4 and 14-6). Suppose there 
is no separate erasing head. If it is desired to write a digit into 
a given cell, the recording head is pulsed in the appropriate 
direction at the time this cell is passing the recording gap. If 
synchronization and pulse shape are correct, the new digit appear- 
ing in the cell will be independent of what was previously stored 
in the cell. To read a given cell, the output of the reading head 
(which may also be the recording head) is gated through at the 
time the desired cell is passing the gap of this head. Associated 
with each discrete digit position around the periphery of the 
drum is a number, the binary representation of which is per- 
manently recorded on the drum in a group of address tracks. A 
given cell is selected by referring to these address numbers by 
means of coincidence circuits. 

This method, which may be termed selective mark insertion, 
entails the considerable problem of entering a digit into a cell in 
such a precise manner that the previously stored digit is com- 
pletely eradicated. The principal advantage of the method 
over continuous erasure is that the stored information is fixed, 
or nonvolatile; i.e.; failure of power or of the writing system does 
not destroy information stored on the drum. In fact, magnetic 
recording in this fashion appears to be the only means of storage 
which is both erasable and nonvolatile and at the present time 
capable of reasonable speed at reasonable cost. 



TRANSFER MEDIUMS 333 

Another method of transferring data to and from the drum is to 
apply the continuous erase-rewrite recirculating techniques similar 
to those used in acoustic-delay-line storage systems. 44 - 45 The 
carrier travels successively past the reading head, an erasing 
head which continuously removes previously recorded data, 
and the recording head. Signals from the reading head are 
reshaped and fed to the recording head. By means of suitable 
interposed gates, it is possible to enter a new digit by impressing 
it on the recording-head circuit at the proper time in place of the 
recirculating digit coming from the reading head. Recorded 
digits may be selectively read out by gating the reading-circuit 
output into the output storage register at the appropriate time. 
Since any delay in the rewriting circuit, even if introduced 
deliberately, is apt to be considerably shorter than the transit 
time of the carrier between the reading and recording heads, 
the stored digital information is repeatedly transferred to new 
positions on the track, and the entire pattern keeps shifting about 
the drum. Location can be handled by a counting track as in 
the spot insertion technique, but it is evidently necessary to 
erase and rewrite the counting track in the same manner as the 
digital tracks. A second control track containing permanently 
recorded timing pulses may serve as a source of clock pulses for 
synchronizing the shaping, gating, and rewriting circuits with 
the drum rotation. Here, as in delay-line storage systems, the 
stored information is volatile if the erasing is done with per- 
manent magnets. 

Track Schemes. Magnetic-drum storage is readily applicable 
to parallel bus computer systems, with one or more tracks on 
the drum serving each place or column of a multidigit number. 
The locating tracks and circuits can be common to all columns. 
If one track is used per column, then the system can store as 
many numbers as there are digits per track. However, the 
location of a given cell for reading or writing may require a 
waiting time of one whole revolution of the drum. The access 
time may be reduced in several ways. One method is to use 
more tracks per column, with a drum of smaller diameter. This 
may be uneconomical of circuit components. A more economical 
method of reducing the waiting time is to install several reading 
heads spaced around the circumference of each track. This, of 
course, helps only the reading problem and does not reduce the 



334 PHYSICAL COMPONENTS AND METHODS 

time to reach a given cell for writing. Installation of a number 
of dual-purpose heads around each track appears feasible, but 
the circuit economy is again questionable. 

Magnetic Design Considerations. In any magnetic storage 
system the speed of operation and over-all size are largely deter- 
mined by the physical and magnetic properties of the magnetic 
medium and the head structures. 

The principal properties desirable in a magnetic recording 
medium and in the head structures for rapid and compact digital 
storage will be described briefly. These properties as applied to 
magnetic recording in general have been discussed in a number of 
papers. 21 ' 39 

The value of the maximum energy product in a medium 
(BH^) should be large. This quantity is indicative of the 
energy which may be stored in a given sized volume of the 
medium and is, therefore, an index of the relative signal strength 
obtainable from the record when wavelengths are short, as 
required in the practical case. This statement is explained by 
the following considerations: 

The coercive force H c must be large to enable a given element 
of the medium to be magnetized independently of adjacent ele- 
ments. High coercive force makes possible improved definition, 
or resolution of the recorded pattern into smaller elements. 

However, if a magnetized element is shortened in the direction 
of magnetization, its demagnetizing factor becomes larger, and 
the residual field available for generation of signals in the reading 
head diminishes. This variation of residual field with length of 
magnetized area is less if the ratio of coercive force to remanence, 
H c /B r , is large. The role of self-demagnetization in the magnetic 
recording process has been treated by Wooldridge, 39 Camras, 21 
and Kornei. 28 

Since the voltage generated in the reading coil is proportional 
to B r , both B r and H c are as large as possible in the best recording 
mediums. 

If the recording signal contains high-frequency components, 
the depth of penetration of the signal into the medium is limited 
by skin effects. If the electrical conductivity of the medium is 
low, the depth of penetration is not so severely limited. This 
is an important advantage of the magnetic oxide coatings over 
solid metallic mediums. If the magnetic coating is kept suffi- 



TRANSFER MEDIUMS 335 

ciently thin, there will be no great differences in effective depth, 
or thickness, of recorded signals at different frequencies, and 
consequently no great difference in flux available for generation 
of signal voltage. If a-c erasure is contemplated, these same 
factors must be recognized in order that the erasing field may be 
capable of penetrating to the deepest level of recorded signal. 

In noncontact longitudinal recording, the fringing flux in 
the vicinity of the gap of the ring-shaped head penetrates the 
carrier and records the pattern of information. The geometry 
of the core and the gap must be carefully chosen so that the 
fringing field exhibits the desired properties. The longitudinal 
component of this field must be as large as possible at the inter- 
section of the equatorial plane of the gap and the surface of the 
carrier, at the specified head-to-carrier clearance. For resolving 
power, the flux density must drop off rapidly on either side of the 
equatorial plane. Both the longitudinal extent of the flux pattern 
and its maximum intensity must be as insensitive as possible to 
variations in head-to-carrier clearance. 

The magnetic properties of the head cores must be similar in 
many respects to those of communication-type transformers 
capable of passing wide bands at high frequencies. The initial 
permeability should be high in reading heads, because of the very 
small magnetomotive forces involved. For recording it is 
important that the maximum permeability be high enough to 
prevent saturation. The coercive force should be small in both 
cases to eliminate residual magnetization effects. Alloys such 
as 4-79 Permalloy, Supermalloy, and Mumetal are representa- 
tive. 19 Cores must be built up of suitably insulated thin lamina- 
tions for proper high-frequency response. 

The recording-head winding must supply a large number of 
ampere turns. At the same time, the ratio of inductance to 
resistance in the circuit must be kept low for rapid response. 
This indicates the desirability of a relatively small number of 
turns of large wire, operated at high currents. In the reading 
head, large output voltage is the critical quantity, and therefore 
a large number of turns is indicated. In both types of winding, 
resonant frequencies within the pass band of the system must be 
avoided. 

Ring-shaped heads are frequently made up of two equal halves, 
for ease of fabrication. The additional air gap (see (r 2 , dotted in 



336 PHYSICAL COMPONENTS AND METHODS 

Fig. 14-2d) appearing in the magnetic circuit tends to reduce the 
inductance, the flux due to a given magnetomotive force, and the 
remanent induction. 

Description of a Typical System. Conservative magnetic- 
drum design constants on which ERA has standardized are the 
following: 23 

Digit spacing: 80 digits per inch, on each track around the 

drum. 
Track spacing: eight tracks per inch along the drum surface, 

in the direction parallel to the axis. 
Surface speed: 1,600 inches per second. 
Maximum diameter: 34 inches. 

These same specifications were used in the magnetic-drum exam- 
ple of Sec. 14-1-1 ; reliable storage systems conforming with them 
have been developed and tested. It is probable that reliable 
systems for the storage of from a few thousand binary digits to 
2 million or more on a single drum can be constructed. 

The most suitable recording medium 25 known to the authors at 
the time of this writing is magnetic iron oxide applied in sus- 
pension with a spray gun and protected by a thin layer of hard 
lacquer. A clearance of 0.002 inch is maintained between the 
heads and the drum surface. 

Typically, numbers are stored on the drum in parallel repre- 
sentation; i.e., every digit of a number appears in a separate 
track, and all digits of the same number pass under their respec- 
tive heads simultaneously. 

As another example, suppose we wish to store 10,000 numbers 
ranging from 000,000,000 to 999,999,999. In the binary system 
we need 30 digits to express the number 999,999,999 because 
2 30 = 1,073,741,824. Therefore we need the equivalent of 30 
parallel tracks around a drum, each track long enough to contain 
10,000 digits. (The corresponding considerations pertinent to a 
binary-coded decimal representation may be derived from the 
rules set forth in Sec. 13-4-5.) For the sake of ease of reference, 
as will become clear shortly, it is convenient to design a drum so 
that the number of digits in each track around it is a power of 2, 
though this is by no means necessary. In this example, we shall 
assume a drum containing 8,192 (i.e., 2 13 ) binary digits per track. 
10,000 is between 2 13 and 2 14 , so we need two groups, each of 30 



TRANSFER MEDIUMS 337 

parallel tracks, around the drum to accommodate the 10,000 
numbers. 

In order to locate any number in storage, as was explained in 
Sec. 5-4-2, there must be associated with it an address which 
specifies uniquely the location in the storage medium where that 
number may be found. We need 8,192 addresses to identify the 
8,192 angular positions around the drum and two addresses to 
identify the two 30-track groups. Provision must be made on 
the drum for the addresses of the 8,192 angular positions expressed 
in binary digits; 13 channels will be required. Finally, one 
channel is required for timing. So the drum must have 74 tracks, 
with each of which will be associated one magnetic head. 

Such a drum, which is actually capable of storing 16,384 (i.e., 
2 14 ) numbers of 30 binary digits each — though our assumed 
requirement specified only 10,000 — will be about 34 inches in 
diameter and 10 inches long. The storage cycle time will be 
about 64 milliseconds. About 660 vacuum tubes will be required 
in the associated reading and writing circuits, or about 1.3 
tubes per 1,000 digits stored. This example was chosen to show 
the economy of magnetic-drum storage for large quantities of 
data. In terms of tubes per 1,000 digits, the magnetic drum is 
more economical than the mercury delay^ line, as may be noted 
by comparing the figure given above with the figure of 11 tubes 
per 1,000 digits for a particular mercury delay line machine 
mentioned in Sec. 14-1-1. Another major advantage of the 
drum is nonvolatility of the recording. On the other hand, the 
storage cycle period is relatively long. As was pointed out in 
Sec. 12-2, the choice of a particular machine must be made 
with all the problem requirements in mind. 

Magnetic drums can be built to store many more or fewer 
digits than the drum postulated above. Table 14-1, a con- 
densation of a table prepared by ERA, 23 shows a few possible 
combinations of design parameters and the corresponding cost 
of each system in tubes per 1,000 digits. 

Reliability. The life expectancy of magnetic wire and tape 
depends upon whether or not the heads are placed in contact 
with the medium. Home recording equipment built by the 
Brush Development Co. employs coated paper tape and contact 
heads, and the tape is specified to be good for 1,000 playings. 
Plastic tape and plated tape should wear better than coated tape, 



338 



PHYSICAL COMPONENTS AND METHODS 



and if the heads are not placed in contact with the medium, its 
life is limited only by the wear resulting from passing it from one 
reel to another. 









Table 


14-1 










Storage capacity 


Storage 
cycle 
time, 
milli- 
seconds 


Drum dimen- 
sions, inches 


Num- 
ber of 
mag- 
netic 
heads 


Num- 
ber of 
track 
groups 


Num- 
ber of 

tubes 






Corre- 
sponding 
number 

of 30- 

digit 
numbers 


Tubes 

per 

1,000 

digits 


Number of 
binary 
digits 


Diam- 
eter 


Length 


61,440 


2,048 


8 


4.3 


10 


71 


2 


630 


10 


122,880 


4,096 


8 
16 


4.3 

8.5 


18 
10 


131 

72 


4 
2 


820 
640 


6.7 
5.2 


245,760 


8,192 


8 
16 
32 


4.3 

8.5 
17 


33 

18 
10 


251 

132 

73 


8 
4 
2 


1,200 
830 
650 


4.9 
3.4 
2.6 


491,520 


16,384 


16 
32 

64 


8.5 
17 
34 


33 

18 
10 


252 
133 

74 


8 
4 

2 


1,210 
840 
660 


2.5 

1.7 
1.3 


983,040 


32,768 


32 

64 


17 
34 


33 

18 


253 
134 


8 
4 


1,230 
850 


1.3 

0.87 


1,966,080 


65,536 


64 


34 


33 


254 


8 


1,240 


0.63 



Magnetic storage mediums are reliable and do not wear rapidly 
when noncontact heads are used. The record is nonvolatile 
and insensitive to temperature, humidity, and small amounts of 
scratching or dust. 

When the reading or recording heads are maintained in contact 
with, the medium, there is a tendency to scrape up magnetic 
material and fill the air gap. The reliability under such circum- 
stances therefore is rather unpredictable. 

The signal level at the output or reading head is low and 
requires voltage amplification in order to obtain a reasonable 
pulse size. The impedance of a writing head is low, so power 
amplification is required at the input to the system. 



TRANSFER MEDIUMS 339 

Size and Cost. The capacity of magnetic wire and of coated 
paper tape is in the order of 10 10 digits per cubic foot and 3 X 10 9 
digits per cubic foot, respectively. The cost of storing digits in 
the two mediums is about 2.5 X 10 6 digits per dollar on wire and 
10 7 digits per dollar on tape. Read-write speeds are about 10 4 
digits per second for wire and 10 5 digits per second for coated 
paper tape. 

To make supplementary copies of a magnetic wire or tape 
record, it is feasible to connect a reading unit of a computer to 
the recording unit and run them off. Magnetic reading equip- 
ment is required to inspect the record, but checking equipment 
should not prove a particularly expensive addition to a large 
computer. An average of about six tube envelopes is required to 
read from and write on each channel of digital magnetic record. 

Conclusions. Magnetic mediums are particularly suitable in 
situations where alterations of the recorded data are desirable 
from time to time, since new data may be recorded over old. 
It is also convenient to be able to correct errors at the time that 
records are being made. 

A double-ended medium is easier to handle than a loop, and 
tape is somewhat easier to handle than wire. Tape is more 
versatile than wire since only longitudinal recording can be 
used on the latter. 

Magnetic materials are also considered an excellent medium for 
a permanent storage system such as might be used at the input 
and output of a digital computer. Two possible configurations 
are of the most interest: double-ended wire or tape and loops of 
wire or tape. The former is most convenient for data which are 
to be read in or out only once, and the latter is useful for tables 
or other data which must be referred to frequently and possibly 
in random order. 

14-2-5. Phosphor Storage. The persistence of light radiation 
from phosphorescent materials has been considered as the basis 
for digital storage; O'Neal and Tyler of Eastman Kodak Com- 
pany 11 and Moore of Harvard 41 have reported their early findings 
on this subject. A light-spot pattern may be placed on a phos- 
phorescent medium at a given time and place and be detected 
at a later time and different place. Since phosphorescent storage 
of light is volatile, the light spots must be regenerated periodi- 
cally. Therefore, phosphors are considered suitable as cyclic 



340 PHYSICAL COMPONENTS AND METHODS 

dynamic storage mediums. They are best used when in the 
form of a coating upon a rotating disk or drum similar to the 
magnetic drum described in the preceding section. 

When a phosphorescent material is energized by electromag- 
netic radiation, it continues to radiate energy, usually of a longer 
wavelength, for some time after the excitation energy has been 
removed. The decay of energy from a phosphor is exponential 
in character, and therefore the radiated energy must be utilized 
before the elapse of a time corresponding to a certain minimum 
value of radiation. It is also desirable to erase the record com- 
pletely before recording on the same area again. This is to 
ensure a uniform degree of discrimination between areas in an 
excited state and areas in a nonexcited state. A method of 
erasing is to irradiate the phosphor with infrared. Experiments 
have shown that thorough erasure by this procedure may be 
obtained only at the expense of overheating the carrier medium. 

Phosphors were developed during the Second World War 
which have the property that they store a considerable amount of 
energy when irradiated with ultraviolet light and release visible 
light when stimulated by infrared. Further irradiation by 
infrared completely removes the stored energy. The process is 
cyclic, and the phosphor is not consumed in being returned to 
its original energy-free state. These phosphors are being studied 
at the Eastman Kodak Company. 

Because the energy level of a phosphor decreases with time, 
only two states, energized and nonenergized, are readily usable in 
a phosphor storage device. An important factor in determining 
the usefulness of phosphors as a storage medium is the resolu- 
tion which can be obtained. From this factor, the capacity and 
access time of a drum of reasonable size can be computed. How- 
ever, this information has not been announced. 

The writing equipment for a phosphor storage system consists 
of a visible or an ultraviolet light source which is capable of being 
modulated at a high frequency. The light source is projected 
as a small spot on the periphery of a phosphor-coated disk or 
drum. Electrical pulses corresponding to digits of information 
are supplied to modulate the light source as the disk or drum is 
rotated. Energized spots are then produced sequentially around 
the periphery to correspond to the pulsations of light. 

The reading equipment includes an optical system to project 



TRANSFER MEDIUMS 341 

on a photocell the light from an area on the drum surface. If the 
phosphor used must first be stimulated by infrared radiation, a 
source of infrared is located ahead of the phototube. 

The reliability of a phosphor storage system is difficult to 
ascertain. There are engineering problems of light shielding, 
heat dissipation from the system, and dust elimination from the 
optical equipment. On the other hand, the spacing between read- 
ing and writing heads and the surface of the drum is probably not 
as critical for the phosphor drum as it is for the magnetic drum. 

At the present time, the phosphor storage system does not 
seem to compete seriously with other better developed devices. 

14-2-6. Sonic-delay-line Storage. One of the most adaptable 
means of high-speed dynamic storage is the sonic delay line. 42-50 
Binary digits have been stored in these lines with a digit duration 
in the order of microseconds. Sonic delay lines have been 
selected as the storage medium and as part of the arithmetic 
unit of several large computing machines, including the EDVAC, 
the BINAC, and the proposed UNIVAC. The first of these 
machines was built by the Moore School for the U.S. Army 
Ordnance Department; the second was built by the Eckert- 
Mauchly Computer Corporation for Northrop Aircraft, Incor- 
porated ; and the third is undergoing final development and con- 
struction at the Eckert-Mauchly Corporation at the time of this 
writing. 

Mediums Available and Modes of Transmission. Sonic delay 
lines function by transmitting sound pulses (usually at ultra- 
sonic frequencies) through a chosen medium. The general nature 
of this medium is of prime importance and deserves first con- 
sideration in a description of a sonic storage system. Gaseous 
mediums seem not particularly useful for this application because 
of high attenuations and other obvious difficulties. Either solid 
or liquid mediums seem to offer some promise, although at the 
moment liquid mediums seem more convenient. This con- 
venience arises from the fact that in liquids only one type of wave 
motion is important, the compressional mode. Since there can 
be no elastic shearing force in liquid, the only transverse waves 
possible are surface waves. On the other hand, solids exhibit 
pronounced tendencies to transmit transverse waves as well as 
numerous other modes, and all these modes propagate with differ- 
ent velocities. 



342 PHYSICAL COMPONENTS AND METHODS 

The compressional wave in many liquids propagates with a 
velocity in the neighborhood of 5 feet per millisecond, more or 
less, depending upon the medium. In particular, in mercury, 
which is the most popular medium for delay lines for computing 
devices, the propagation time is just slightly less than 17.5 
microseconds per inch at normal room temperature. The delay 
time for compressional waves in quartz is about G.3 microseconds 
per inch. Compressional waves are likely to be propagated 
one and a half to two times as rapidly as transverse waves, 
depending upon the elastic properties of the substance. In 
fused quartz, for example, the velocity of the transverse wave is 
about 63 per cent of the velocity of the compressional wave, 
so that the delay time is about 10 microseconds per inch. 

Four mediums have been seriously considered for delay lines. 
These are fused quartz and magnesium alloys for solid delay 
lines, and mercury and aqueous solutions of ethylene glycol or 
other solutes for liquid delay lines. The transducer for convert- 
ing sonic to electrical energy is a piezoelectric crystal, usually 
quartz. 

The choice of the medium to be used is usually based upon one 
or more of the following factors, the importance of which must 
be estimated in connection with the particular problem at hand: 

1. Required delay. 

2. Allowable duration of digits. 

3. Attenuation, and minimum allowable signal-to-noise ratio. 

4. Discrimination against spurious signals and echoes. 

5. Dependence of delay on temperature, pressure, frequency, 
etc. 

6. Cost, availability of materials, ease of manufacturing, etc. 

7. Reliability, ruggedness, etc. 

General Structure of Delay Lines. The general structure of 
delay lines is described in the literature. 42,43 The structure 
usually consists of two quartz-crystal transducers separated by 
the transmitting medium. The resonant frequency of these 
transducers may be from 5 to 30 megacycles. One is a trans- 
mitting transducer; it receives electrical signals and through 
piezoelectric contractions transforms them to sonic signals in 
the medium. The other is a receiving crystal ; it vibrates under 
the influence of the sonic signals from the medium and generates 



TRANSFER MEDIUMS 343 

electrical signals. Both crystals are cut with orientation of 
anisotropic axes so that the mode of vibration will introduce 
waves of the type chosen — compressional waves in liquid and, 
possibly, transverse waves in solids. In order to gain good 
directional properties of transmission, the crystal faces are plane 
and several wavelengths in dimension. The crystals are care- 
fully aligned so that the faces of the transmitting and receiving 
crystals are as nearly parallel as they can be made, and so that 
the perpendicular from the center of one face strikes the center 
of the other. In some cases reflectors are used to reduce the 
over-all length of the delay lines while maintaining a given 
acoustic length. The amount of delay depends upon the 
velocity of sound waves through the medium and upon the 
distance between the transducers. 

In the past, for computing purposes, binary digits have been 
used; a burst of signal persisting a fraction of a microsecond is 
used to denote the digit 1, and no signal at all is used to denote 
the digit 0. An alternate scheme might be to use two different 
frequencies to represent the two digits and 1, or r different 
frequencies to represent different digits radix r. 

Acoustical Match. The choice of mercury as the medium 
for sound transmission in liquid delay lines is dictated mainly 
by consideration of acoustical impedance. Where sound is 
being transmitted across a boundary from one substance into 
another, reflection will normally take place. The fraction of 
energy reflected and the fraction transmitted through the sub- 
stance as a ray strikes the boundary perpendicularly depend 
upon the degree to which the acoustical impedances of the two 
substances match. Acoustical impedance is a parameter which 
depends upon the elastic properties of the substances. The 
only known liquid with an acoustical impedance comparable to 
that of quartz is mercury. This implies that the vibrational 
energy of the transmitting transducer is transferred efficiently 
to the mercury and that most of the energy transmitted to the 
receiving transducer is absorbed by this transducer and converted 
to electrical energy. 

The efficiency of transfer of energy is important in several 
ways. In the first place, there is the obvious desire to transmit 
strong signals through the medium; this, however, is probably 
not as important a consideration as two others. The energy 



344 PHYSICAL COMPONENTS AND METHODS 

not absorbed by the receiving crystal is reflected back to the 
transmitting crystal, and it is again reflected by the transmitting 
crystal to the receiving crystal; thus, if there is not a good 
impedance match between the transducer and the medium, 
the possibilities of echoes are enhanced, and signals may be 
modified seriously by these echoes. Finally, transducers nor- 
mally are very sharply resonant, and some steps must be taken 
to decrease the Q of the crystals in order to increase the obtainable 
bandwidth. The most efficient way of decreasing the Q is by 
loading the crystal mechanically with the transmitting medium 
itself ; in this way the Q is decreased by virtue of the fact that the 
energy is transferred at a high efficiency to the medium. 

Fused quartz is a popular choice for solid delay lines because 
of the absence of many objectionable features found in other 
solids. Its acoustical match to quartz crystals is extremely 
good. Of the other two types of delay lines already mentioned, 
the match between magnesium alloys and quartz is so poor, and 
that between aqueous solutions and quartz is so poor, as to make 
these delay lines probably useless in computing machine applica- 
tions. Magnesium alloys, however, have a somewhat better 
acoustical match with ADP* crystals, and the match between 
the solid delay lines and the quartz crystals can conceivably be 
improved by judicious choice of cement between the crystal and 
the transmission medium. 

Temperature Variations. The velocity of sound in mercury 
at frequencies of a few megacycles will change about 1 part 
in 3,000 for each degree centigrade change in temperature. 
This means that if 3,000 binary digits are stored in the delay 
line, and if no log is kept of the digits emerging, and if the tem- 
perature in the delay line is not known to within 1°C, then even 
though the individual digits emerging can be read perfectly, 
it is impossible to tell which digit is being read at a given time. 
This implies practical limitations on the length of mercury delay 
lines. Other substances show a similar dependence of velocity 
on temperature. Fused quartz, used with transverse waves, 
has a temperature dependence less than half that of mercury; 
this implies that more than twice as many digits can be stored in 
quartz as in mercury before temperature troubles arise, assuming 
that the temperature control of each substance is comparable. 

* Ammonium dihydrogen phosphate, NH4H2PO4. 



TRANSFER MEDIUMS 345 

This assumption is probably not completely justified. Pure 
water shows a zero dependence of velocity on temperature for 
temperatures around 74°C. For lower temperatures, the 
dependence is fairly marked ; however, some solutes, in particular 
alcohol and ethylene glycol, can be used to reduce the temperature 
at which this zero temperature coefficient occurs. This is the 
consideration which has led to the adoption of aqueous solutions 
in sonic delay lines. However, in the case of most computing 
applications, where the pulse repetition rate is so great that 
echoes could cause difficulty, this temperature consideration is 
overruled by the interference consideration. 

Attenuation. The theory of transmission of sound through 
liquids predicts an attenuation of sonic energy proportional 
to the square of the frequency. This condition is confirmed for 
frequencies of the order of a megacycle. For mercury, attenua- 
tion is about 5 decibels per millisecond at 10 megacycles. It is 
therefore not a serious consideration at this frequency and 
reported results of propagation measurements indicate no sig- 
nificant change in velocity or frequency-free absorption coeffi- 
cient with frequency. 48 In solids, the attenuation is nearly 
proportional to the frequency. At 10 megacycles the attenua- 
tions in fused quartz and in mercury are comparable. At 30 
megacycles, fused quartz should still be usable without serious 
attenuation. 

Beam Characteristics of Crystals. For frequencies so high 
that the wavelength of the transmitted sound is short compared 
with crystal dimensions, plane crystals radiate narrow, highly 
directive beams. The width of the beam is governed by the 
usual equations of transmission of plane waves through windows. 
In particular, with crystals of reasonable size and with trans- 
mitted frequencies of a few megacycles, the beam width is suffi- 
ciently narrow to require careful aiming to give efficient trans- 
mission of sound through the medium. For transmission through 
solids, where the wavelength is longer, this directivity is reduced. 
This is one consideration which is of some importance in deciding 
between liquid and solid delay lines. 

Pulse Duration. The shortest pulse which can be successfully 
transmitted is a function of several variables. In the first place, 
it must be realized that the pulse is a modulation of a carrier 
and that, as such, it contains essential side bands with band- 



346 



PHYSICAL COMPONENTS AND METHODS 



Table 14-2. Summary of Important Sonic Properties of Delay-line 

Mediums 





Magnesium 


Fused 


Mercury 


Alcohol 


Property- 


alloy 


quartz 


solution* 




(Solid) 


(Solid) 


(Liquid) 


(Liquid) 


Match with 


Fair (with 


Excellent 


Good 


Unsatisfactory 


crystal 


ADP) 








Dependence of ve- 


About 1 


About 1 part 


About 1 part 


Zero at normal 


locity on tem- 


part in 


in 9,000 per 


in 3,000 per 


temperature 


perature 


2,500 per 


degree cen- 


degree cen- 






degree 


tigrade 


tigrade 






centigrade 








Dependence of ve- 




locity on fre- 




quency 


Too small to be noted ** 


Dependence of at- 










tenuation on 










frequency 


Unknown 


Linear 


Quadratic 


Quadratic 


Minimum binary- 






digit duration 


Less than 5 X 10 -7 second 


Unknown 


Velocity- 


Trans- 


Transverse : 


145 centime- 






verse : 


380 centi- 


t e r s per 






300 centi- 


meters per 


millisecond 






m e t ers 


millisecond 








per milli- 


Longitudi- 








second 


nal : 590 








Longitudi- 


centimeters 








nal: 550 


per millisec- 








t o 580 


ond 








centi- 










meters 










per milli- 










second 








Attenuation 


10 decibels 


10 decibels 


5 decibels per 






per milli- 


per milli- 


millisecond 






second at 


second at 


at 10 mega- 






10 mega- 


15 megacy- 


cycles 






cycles 


cles 







* Other solutions also give zero temperature dependence with poor 
transducer match. 



TRANSFER MEDIUMS 



347 



width depending on the pulse shape and length. The sharply- 
resonant properties of the crystal transducer limit the bandwidth 
which it will accept ; in order to widen this bandwidth, loading of 
the crystal is necessary. It has been pointed out that this 
loading is most effectively brought about by a good match 
between the crystal and the medium. In fused quartz, it seems 
likely that a 6J^-megacycle bandwidth can be attained easily 
for a 30-megacycle carrier. This might permit pulses spaced 
6 to the microsecond. Four pulses per microsecond have already 
been attained in mercury; at least 2 per microsecond have been 



CLOCK 
PULSES 



CONTROLS 




READ 
IN 



Fig. 14-7. Acoustic-delay-line storage system. 



attained in solid delay lines. Roughly, the same consideration 
will hold for either frequency-shift or on-off indications of digits. 

Complete Storage System. The read-write equipment asso- 
ciated with an acoustic-delay-line type of storage system consists 
of a driver which excites the input crystal, an amplifier at the 
output to compensate for the attenuation through the delay 
line, and a set of gates with which to control the course of the 
stream of pulses. These components are illustrated in Fig. 14-7. 
Pulses may be introduced into the tank when G± is open and will 
continue to circulate until G2 is opened, whereupon the stream 
will be terminated at G 2 . All the circulating pulses or any 
portion thereof may be read by opening G$ when the desired 
pulses are leaving the delay tank. 

The components described require about 11 tubes, and the 



348 PHYSICAL COMPONENTS AND METHODS 

complexity of the equipment is independent of pulse-stream 
length. On the other hand, the degree of temperature control 
required does depend upon the pulse-stream length. In order 
to keep the pulses emanating from the delay line synchronized 
with clock pulses, the temperature of the medium must be con- 
trolled closely so that the actual delay time remains nearly 
constant. Another approach to the problem is to mount the 
delay lines together so that they all have the same temperature 
within an allowable tolerance and to use one of the delay tanks to 
control the spacing of the clock pulses. 50 Thus, the delay in a 
tank would correspond to a constant number of pulses even 
though the actual delay time varied with temperature. 

In the BINAC there are 18 separate delay lines or channels in 
the same tank of mercury. This is a cylindrical tank with cir- 
cular cross section about 3 inches in diameter. The capacity 
of each channel is 1,344 pulse positions, including 960 digits 
and the algebraic signs and spaces between digit groups. At 
the effective carrier frequency used in the BINAC (about 12 
megacycles) a transducer of the order of % inch across becomes 
such a sharp-beamed acoustic radiator that no shielding in the 
tank is actually necessary to isolate each beam through the 
mercury from the others, although shielding has been used as an 
additional precaution. 

The possibility of adding reading heads along the path of a 
stream of pulses within an acoustic delay line has been considered, 
but no results are known to have been announced. The attenua- 
tion introduced by such units is large if the acoustic match is 
good ; on the other hand, if the acoustic match is poor, there will 
be echoes in the tank and the true signals will be difficult to 
distinguish. 

Conclusions. To date, more work has been done on mercury 
delay lines than on other types. The cost of such systems of 
storage is estimated to be about $1 for 10 binary digits, and the 
space required is approximately 1 cubic foot for 2 X 10 4 binary 
digits. 

Mercury delay lines have been used for some time in several 
applications, including radar equipment and training devices. 
They are reasonably reliable. Their only serious defect is the 
variation of velocity with temperature, but this may be handled 
in a straightforward manner. 



TRANSFER MEDIUMS 349 

14-2-7. Electromagnetic Transmission-line Storage. An obvi- 
ous way in which to delay and store electrical pulses is to intro- 
duce them directly into one end of a transmission line and to 
read them at a later time as they emerge from the other end. 51-59 
Transmission lines may be used as cyclic or noncyclic storage 
mediums by regenerating and recirculating the pulses at the 
receiving end of the line or by using the pulses as they emerge 
after a fixed delay time. 

General. The major difficulty with transmission-line storage 
is that an unreasonable length of line is required to store a large 
number of digits. This is true because the phase velocity of 
electrical pulses along ordinary transmission lines is of the order 
of magnitude of the velocity of light. Since the lines must 
necessarily be long in order to be of use as storage mediums, con- 
siderable power is dissipated in the resistance of the conductors 
in the form of ohmic, eddy-current, and hysteresis losses. A 
number of attempts, therefore, have been made to decrease the 
phase velocity along various types of transmission lines in order 
to make them more suitable for use as storage mediums. 

Several features of the pass band are important if a short pulse 
is to be passed through a transmission line and remain distinguish- 
able at the far end. If a pulse has a width p, the line must pass 
at least the fundamental frequency 1/p. Kallman shows 54 that 
nonlinear phase shift of the major frequency components, or 
phase distortion, of the pulse is more detrimental than amplitude 
attenuation, or frequency distortion. Phase distortion of com- 
ponent frequencies tends to spread out the pulse in time and 
causes it to be indistinct. In some situations where a lengthened 
pulse is desired, this property of a delay circuit can be used to 
advantage, but when delay lines are used for pulse storage, 
phase distortion must be held within rather close limits in order 
to get a good pulse-packing factor. 

It is desirable that the pass band of a delay line include the 
pulse-repetition-rate frequencies so that d-c pulses may be passed 
directly. Waveguides, however, cut off low frequencies but 
may be used to store a pulsed carrier which lies within the pass 
band. As a result, the auxiliary equipment necessary for use 
with waveguide storage is more complicated than for wire 
lines. 

Some thought has been given to the possibility of storing digits 



350 PHYSICAL COMPONENTS AND METHODS 

of radix higher than 2 by representing each with a corresponding 
frequency. This possibility is described in more detail in Sec. 
1G-2, but it might be noted that a wide-band transmission line 
would be suitable for the storage of these frequency-shift pulses. 

Three familiar forms of transmission line have been considered 
for computer storage. They are the waveguide, the wire trans- 
mission line, and the artificial line with lumped parameters. 

Waveguide Storage. If the techniques were available to utilize 
pulse repetition rates much greater than 10 6 pulses per second, 
advantage might be taken of the bandwidth of waveguides since 
it is much greater than that of other types of transmission lines. 

A practical limitation on the bandwidth of frequencies which 
can be sustained in a waveguide is determined by the disturbances 
to the structure caused by the input and output coupling mecha- 
nisms. By careful design of the couplers, it is possible to make 
the standing-wave ratio caused by the coupling less than 1.15 
over a frequency deviation range of 10 to 15 per cent. As an 
example, if a carrier frequency of 3,000 megacycles is used, and a 
standing-wave ratio of less than 1.15 must be maintained, a 
maximum bandwidth of 900 megacycles is possible. 

A waveguide with a bandwidth of 900 megacycles may be used 
to store dbuble-side-band pulsed carriers at a repetition rate in 
the order of 10 8 pulses per second, which is 100 times as great 
as for the example cited using a sonic delay line. However, the 
capacity of the waveguide is only 1 digit for each 10-foot length, 
and the attenuation problem in a waveguide long enough to 
store 10 3 digits is severe. In a 1.5- by 3-inch waveguide the 
attenuation is approximately 1 decibel per 100 feet, so the total 
attenuation for the storage of 10 3 digits is about 100 decibels. 
Amplification over a bandwidth of 900 megacycles to restore the 
signals to their original amplitude is not feasible with conven- 
tional tubes, but it is possible with traveling-wave tubes. 

Some work has already been done on low-velocity waveguides. 
For example, in the linear accelerator being built at Stanford 
University, waves are slowed down in a waveguide by means of a 
series of irises placed inside the guide. Although equidistant 
irises are interesting, in that they are effective in slowing down a 
wave, they cause the guide to act like a band-pass filter so that 
attenuation and time delay vary considerably with frequency. 
This results in pulse distortion; hence a low-velocity waveguide 



TRANSFER MEDIUMS 351 

of this type is of limited usefulness as a wide-band storage 
medium. 

Transmission Lines with Distributed Constants. A considerable 
amount of work has been done on wire transmission lines to 
reduce phase velocity. As a result, solenoidal lines have been 
developed for use in equipments including synchroscopes and 
pulse-forming networks. 5 1 • 53, 54, 58 * 59 

In a wire transmission line at relatively low frequencies, if 
losses are neglected, the time required for a pulse to travel along 
a unit length of line is represented by the equation 

T = y/W 

where T = delay time per unit length 

L = series inductance per unit length 

C = shunt capacity per unit length 
The first attempts to increase the delay time resulted in a line 
with one wire wound in the form of a solenoid inside of or sur- 
rounding a conductor of large, surface area. The effect of the 
solenoidal winding is to increase the series inductance of the 
line owing to the increased self-inductance and the introduction 
of mutual inductance between turns. The other conductor, 
formed in a semicylindrical shape, increases the capacity from 
one turn to the next. In order that it may not present a shorted 
turn in parallel with the turns of the solenoid, the semicylindrical 
conductor has an insulated longitudinal strip. The resulting 
structure is quite complicated to analyze; however, Rudenberg 
has shown 56 that waves follow the spiral with approximately the 
velocity of light so that the delay time may be represented by 
the equation 

T ^ — 
c 

where T = time delay per unit length of line 

l w = length of wire per unit length of line 
c = velocity of light 
This expression results in a good approximate design formula 
for a spiral delay line at low frequencies. 

When pulses of short duration are transmitted through a delay 
line of the solenoidal type, it is found that the various frequency 
components do not reach the receiving end at the same time, 



352 PHYSICAL COMPONENTS AND METHODS 

owing to phase distortion. This effect is caused in part by the 
high-frequency path through the turn-to-turn shunt capacity 
along the solenoid and in part by a decrease in effective series 
inductance as the wavelengths of frequency components approach 
the dimensions of a turn. The latter effect results from the fact 
that when the phase angle between the fields surrounding 
adjacent turns is reduced, the effective mutual inductance 
between turns is also reduced. 

One of the first low-velocity transmission lines being manu- 
factured is a solenoidal delay line built by Federal Telephone and 
Radio Corp. and reported by Zimmerman. 69 It is the RG-64U 
cable and has a delay of 0.042 microsecond per foot and an 
attenuation of about 0.14 decibel per foot at 5 megacycles. 
Therefore, the attenuation is about 3.3 decibels per microsecond 
delay. However, this line is not equalized and causes a decrease 
in delay of approximately 0.2 per cent per megacycle in the 
5-megacycle range. The number of microsecond pulses which 
can be stored in a line of this type must therefore be limited to 
about 40 pulses to prevent undue phase distortion. This cable 
was designed as a high-impedance cable rather than a delay line; 
hence presumably greater delay can be achieved if desired. 

A spiral delay line reported by Kallmann, 54 in which various 
shapes and sizes of conductors were placed adjacent to the coil in 
order to increase its distributed capacity, has a velocity as low 
as c/4,300; however, this is attained at the expense of high 
attenuation and a narrow bandwidth. It seems reasonable to 
expect that velocities of c/ 1,000 might be obtained with a band- 
width of 10 to 20 megacycles. This corresponds to a delay of 
1 microsecond per foot of line and two to four pulses per foot. 

Kallmann has reported considerable work to equalize the phase 
shift in delay lines by means of strips of conductors, which are 
placed between the spiral winding and the cylindrical conductor 
and insulated from both. When these conductors are arranged 
correctly, they introduce an equalizing effect which reduces the 
variation of phase velocity with frequency over a band of the 
order of 5 megacycles. Other methods of improving the delay 
characteristic, such as increasing the diameter of the coils and 
separating the windings, were also investigated. When the 
windings are separated, the line might be considered to be a 



TRANSFER MEDIUMS 353 

semilumped-parameter line, similar to the lumped-parameter 
lines discussed in the next section. 

Artificial Transmission Lines with Lumped Parameters. Delay 
lines with lumped parameters have been in use for many years 
as low-pass filters and artificial lines for laboratory experiments. 
A convenient form is the iterative pi or T with inductances in 
the series arms and capacity in the shunt legs. Actually the 
distinction is unclear between spiral lines with pieces of foil to 
increase the shunt capacitance and lines with lumped parameters. 
In either case, there is mutual inductance between adjacent turns 
and shunt capacities across various groups of series turns. A 
series of simple pi networks does not have a flat pass band, but 
it has been found that a mutual inductance M between adjacent 
inductance elements L, adjusted so that M = 0.1 2L, will equalize 
the attenuation and phase shift over the pass band. Lumped- 
parameter lines have been built in the laboratory, but they are 
rather difficult to duplicate with manufacturing processes. 
Brillouin 52 estimates that the velocity in a lumped-parameter 
line may be reduced to c/10,000, with a bandwidth of about 
5 megacycles. This corresponds to a delay of about 10 micro- 
seconds per foot or a packing factor of about 20 pulses per foot of 
line. 

Transmission lines with lumped constants or semilumped 
constants may be built with less attenuation per unit delay time 
than the continuous-wire lines which have been produced to date. 
Semilumped lines can be equalized and probably made suitable 
for bandwidths up to 15 or 20 megacycles. Because of their 
attenuation, they too are quite limited with regard to the number 
of digits which they can store. One of Kallmann's best lines 
had a transmission loss of 3 decibels per microsecond at 5 mega- 
cycles. This corresponds to a loss of about 6 decibels per digit 
at 5 megacycles, so approximately one stage of amplification is 
required for every two or three digits stored. 

Conclusions. Not enough use has been made of electromag- 
netic delay lines in computer applications to prove their depend- 
ability, but experience with delay lines in other applications indi- 
cates that they should be quite dependable. In systems pres- 
ently contemplated they are designated for applications in which 
less than 40 to 100 digits are to be stored or where short adjusting 



354 



PHYSICAL COMPONENTS AND METHODS 



delays or pulse shaping are desired. A comparison of the various 
types of delay line discussed above is shown in Table 14-3. 

Electromagnetic transmission delay lines are being manu- 
factured commercially by the Federal Telephone and Radio 
Company, General Electric, and the James Millen Company. 



Table 14-3 



Type of 

transmission 

medium 



1.5- by 3-inch wave- 
guide 

RG-64U cable 

Equalized delay lines 
(estimated) 

Lumped-parameter. . 

Semilumped-param- 
eter 



Maximum 
delay, 
micro- 
seconds 
per foot 



0.001 
0.042 
0.042 



1 
10 



Attenua- 
tion, 
decibels 
per foot 



1.2 

0.14 

0.45 



50 



Pass 
band, 
mega- 
cycles 



450 

5 

10 

10-20 
5 

20 



Maximum 

number of 

pulses 

per foot 



0.20 
0.08 
0.16 

2-4 
20 



Attenua- 
tion, 
decibels 
per pulse 



2.5 



14-2-8. Deflection-type Cathode -ray -tube Storage. One of 

the most promising means for storing digital information is the 
acquisition and retention of a charge by an electric capacitor. 
In connection with television work and other developments, 
methods have been devised for constructing small storage 
capacitors fairly cheaply, compactly, and in such a way that 
they can be charged and discharged rapidly and simply by means 
of a low-inertia electron beam. 67 Storage tubes are now being 
developed by the Naval Research Laboratory, 63,64 by the Servo- 
mechanisms Laboratory at MIT, 60-62 by the University of 
Manchester, England, 65 by RCA 68-70 and by the Raytheon 
Manufacturing Co. There are probably other storage-tube 
developments in progress, for some work was carried out during 
the Second World War on tubes which might have had applica- 
tion to computing equipment, and it is presumed that some of 
this work is continuing. 

The National Bureau of Standards is conducting experiments 
to determine what can be done with standard components, 



TRANSFER MEDIUMS 355 

confining its interests to storage techniques rather than tube 
development. 

Dielectric Storage Principle. Storage in these tubes depends 
upon electric charges placed at various points on a dielectric 
plate. As an example of such a device, a tube shaped like an 
ordinary cathode-ray oscilloscope tube might be visualized. The 
dielectric plate is placed at the large end of the tube. Means, to 
be described later, are furnished whereby regions on the surface 
of this plate may be charged to various potentials under the 
influence of the cathode ray directed to these regions in con- 
junction with various fields introduced at these regions by 
additional elements. Storage of many digits is obtained by 
charging each of the chosen regions, one for each digit, inde- 
pendently of the rest. If a grid with 32 divisions on each side 
is chosen as a basic pattern for the charges, then this tube will 
store 32 squared, or 1,024, independent digits. The radix of 
these digits will depend upon the number of different distinguish- 
able potentials which can be introduced to the dielectric plate. 
In the utilization of the principles in a manner to be described, 
this number is definitely limited to two potentials; for other 
utilizations of the same principle it might be increased. 

Secondary Emission Characteristics. Figure 14-8 shows, as a 
function of target potential, the ratio of the rate at which electrons 
are leaving a target spot on a dielectric sheet to the rate at which 
electrons in a cathode ray are bombarding this spot. This ratio 
takes various values because of capture of electrons by the 
dielectric sheet, or target, and because of secondary emission from 
the target; these are functions of the field in the tube and of the 
energy of the impinging electrons, and these in turn are functions 
of the potential of the target relative to the cathode and other 
tube elements. Consider, then, a cathode-ray gun bombarding a 
secondary emitter at the end of a tube shaped like a conventional 
cathode-ray tube. Near the target end of the tube is inserted 
some sort of collector element which will catch any electrons 
which are drifting in the tube after reflection from the target 
and will also catch the secondary electrons emitted randomly 
from the target. As ordinate of the curve is plotted the ratio of 
the number of electrons received by the collector to the number 
of electrons in the cathode ray (for convenience this is called the 
effective secondary-emission ratio); as abscissa is plotted the 



356 



PHYSICAL COMPONENTS AND METHODS 



potential of the secondary-emitter target relative to the potential 
of the cathode. A qualitative explanation of the shape of this 
curve is as follows : 

When the secondary-emitter target has a potential just higher 
than the cathode potential, electrons arrive with insufficient 
energy to knock out an appreciable number of secondary elec- 
trons, and the main field between the cathode and the secondary 
emitter holds the received electrons at the secondary emitter; 



EFFECTIVE 
SECONDARY 
EMISSION 2 
RATIO 




POTENTIAL OF SECONDARY EMITTER TARGET 
Fig. 14-8. Effective secondary-emission characteristics. The parameter 
plotted as ordinate is the number of electrons collected by the screen per 
primary electron, i.e., per electron in the cathode beam. This ratio is 
plotted as a function of the potential of the target bombarded by the cathode 
ray. The number of electrons collected includes both the number due to 
secondary emission alone and the number due to electron optical effects 
and reflections from the target as well. If 7 2 stands for the secondary- 
emission current, I\ for the primary (cathode-beam) current, and I r for 
the effects of reflection, the ratio plotted is (7 2 + Ir)/I\. This should not 
be confused with the conventional plot of the ratio I2/I1, which is the true 
secondary-emission ratio. 



thus the electrons are captured, and there is a point on the curve 
where the ratio is practically zero just above the cathode poten- 
tial. When the target is lower in potential than the cathode, the 
original cathode beam is turned back by the dielectric and 
captured by the collector, giving an effective secondary-emission 
ratio of 1. As the target voltage increases, the energy of the 
impinging electrons is increased, and they knock out more 
secondary electrons. The maximum of the effective secondary- 
emission ratio may be well above 1; i.e., many more electrons 
may be collected than are received by the target. However, as 
the potential of the target increases further, the energy required 



TRANSFER MEDIUMS 357 

for electrons to escape from it also increases, and the curve may 
fall off. 

By placing a screen close to the front of the target and main- 
taining it at a constant potential, severe and desirable modifica- 
tion of the high-potential end of this curve may be achieved. 
Here the assumed structure will be an electron gun (as before) 
with a comparatively long path to a conductive screen, which 
serves as a collector for scattered electrons and through which 
practically all the beam electrons pass. Fairly close behind this 
screen, the dielectric secondary-emitting target is placed. Fig- 
ures 14-9 and 14-10 show how the curve is modified by this screen 
when it is at different potentials. The screen has little effect on 
the part of the curve corresponding to target potentials much 
below the screen potential, but it profoundly affects the higher 
potential parts. In Fig. 14-9, the screen serves to catch rejected 
electrons in the low-potential part of the curve, but it does not 
affect the energy with which the primary beam impinges on the 
target and, therefore, does not affect the secondary-emission 
characteristics at this point. However, when the potential of 
the target is higher than the screen potential, the steep potential 
gradient between the target and the screen increases the energy 
required for secondary escape from the target, and the secondary- 
emission ratio is reduced practically to zero, as shown. The same 
effect is illustrated in Fig. 14-10 with the screen potential so 
low that the secondary-emission ratio is held at a value below 1 
for all target potentials above the cathode potential. 

The importance of the unity value for the secondary-emission 
ratio is that it separates the region in which the potential of the 
target will increase from the regions in which this potential will 
decrease when the primary beam impinges on it. When this 
ratio is 1, the number of electrons leaving the target equals the 
number arriving, and there will be no change in target potential. 
When this ratio is greater than 1, the number of electrons leaving 
the target is greater than the number arriving ; under this condi- 
tion, the target loses negative charge, and its potential increases. 
When this ratio is less than 1, the number of electrons arriving 
at the target exceeds the number leaving; the target accumulates 
negative charge, and its potential decreases. Thus, in Fig. 14-9, 
the dielectric target retains its potential if this potential is below 
that of the cathode. (Here minor overshoots of the secondary- 



358 



PHYSICAL COMPONENTS AND METHODS 



emission characteristics due to variations in electron energies are 
ignored; they simply shift the potentials being discussed here 
by an insignificant amount, if the dispersion of energies is small.) 
For potentials between the cathode potential and a critical 
potential, at which the curve crosses the unity ratio value in an 
upward direction, the potential of the target will drop. Thus, 
a target potential anywhere between the cathode potential and 
the critical potential will decrease under irradiation by the elec- 
tron beam until it reaches the cathode potential. Similarly, if 
the potential of the target is between the critical value and that 
of the screen, the target will increase to the screen potential 



EFFECTIVE 

SECONDARY 

EMISSION 

RATIO 



CATHODE CRITICAL SCREEN 

POTENTIAL POTENTIAL POTENTIAL 

REGION OF REGION OF 
STATIONARY FALLING 
TARGET ! TARGET 

POTENTIAL POTENTIAL 

/ 




POTENTIAL OF SECONDARY EMITTER TARGET 

Fig. 14-9. Ratio of collected electrons to electrons in cathode ray plotted 
as a function of target potential. Screen potential above critical value. 

under irradiation ; and if the target is above the screen potential, 
it will fall to the potential value of the screen. Thus, there are 
only two stable target potentials not lower than the cathode 
potential for the configuration indicated in Fig. 14-9; one of 
these is the cathode potential, and the other is the screen poten- 
tial. In Fig. 14-10, the only stable potential not less than the 
cathode potential is the cathode potential itself. 

The structure of a tube to fulfill these conditions is such that 
the target cannot long maintain a potential below that of the 
cathode because of leakage. Actually, the screen itself becomes 
a cathode under such conditions, and the effective secondary- 
emission ratio becomes just larger than 1, tending to move the 
target potential toward the cathode potential; however, the 
screen is usually an unproductive cathode, so this effect is not 
shown on the curves. 

Writing Principles. There are many parameters available 
for use to represent digital values by this configuration, and 
various choices have been made by different workers. The effect 



TRANSFER MEDIUMS 359 

desired, as has been stated already, is the ability to charge 
discrete regions of the dielectric target independently to potentials 
chosen from a set of two or more, and to determine which of the 
set of potentials was placed at each region. Inasmuch as the 
target is chosen as a nonconductor, the potentials at the different 
spots will remain essentially unchanged for some while, and it is 
probable that the regeneration of each spot potential will not 
require any substantial fraction of operating time of the tube. 
The parameters available are (1) potential of the dielectric 
target; (2) screen potential; (3) cathode potential. 

CATHODE 
POTENTIAL 



EFFECTIVE 

SECONDARY 

EMISSION 

PATIO 



I I 

REGION OF REGION OF 
STATIONARY FALLING 
TARGET TARGET 

POTENTIAL POTENTIAL 

/ I ^ 
i SCREEN 
POTENTIAL 
1 




POTENTIAL OF SECONDARY EMITTER TARGET 

Fig. 14-10. Ratio of collected electrons to electrons in cathode ray, plotted 
as a function of target potential. Screen potential below critical value. 

One method of utilizing the secondary-emission characteristics 
represented in Fig. 14-9 will now be described. Here the screen 
potential and the cathode potential are held fast. The effects of 
these restrictions are represented by the curve in Fig. 14-9, with 
the cathode potential and the screen potential held constant. 
The potential of the dielectric is the variable; its two stable 
points, one at the cathode potential and one at the screen poten- 
tial, are used to represent the binary digits and 1. One addi- 
tional element is added to the tube to change the potential of the 
dielectric; it may be a capacitor plate placed outside the tube so 
that the secondary-emitting dielectric lies between it and the 
screen. Once one of the stable potentials is attained by the 
dielectric, this potential will be held while the target is being 
irradiated and for a considerable time after irradiation has 
ceased. Thus, if the target is at the cathode potential, it will 
remain at the cathode potential under bombardment by the 
cathode beam. However, the target potential can be increased 
momentarily by introducing a strong positive charge on the 



360 PHYSICAL COMPONENTS AND METHODS 

additional capacitor plate outside the tube. The capacitor 
action between this element and the screen will introduce a poten- 
tial at each point of the dielectric between them, and if the 
potential of the outside plate is made sufficiently high, the 
potential of the point on which the electron beam is impinging 
will rise to a value greater than the critical potential shown in 
Fig. 14-9. If this happens while the secondary emitter is being 
bombarded, the potential of the secondary emitter will drift 
upward to the screen potential, and it will remain there if the 
charge is gradually removed from the outside plate. Thus, 
a charge can be placed on the outside plate to force the potential 
of the irradiated dielectric from the value corresponding to the 
digit to the value corresponding to the digit 1. If the potential 
had had this value in the beginning, the added potential resulting 
from the charge on the outside plate would have been com- 
pensated for by decreased secondary emission as shown in the 
curve; in either case, whether or 1 was originally stored, the 
final result will be that the digit 1 is stored. In the same way, 
a negative charge introduced to the outside plate may be used 
to force the dielectric potential to a value below the critical 
potential, whether the original potential was at the 1 or level. 
When this charge is gradually removed, the potential of the 
irradiated dielectric will stabilize at the cathode potential, and 
the digit will have been stored. There will be no change in 
the information stored at the points of the dielectric which are 
not being bombarded by cathode rays during potential changes 
introduced by the charges placed on the outside plate, for 
although their potential may be forced above or below the critical 
potential, there is no change in charge on them during this period, 
for no electrons impinge on them and no secondary electrons are 
emitted. 

A second scheme for storing and changing data involves varia- 
tions of screen potential. In this method, the only stable point 
used on the secondary-emission curve is the point at the screen 
potential in Fig. 14-9, and care is taken never to permit the 
potential of the secondary emitter to drop below the critical 
potential. Several different potentials can be assumed by the 
secondary emitter corresponding to several different screen 
potentials. Thus, a screen potential can be chosen and a point 



TRANSFER MEDIUMS 361 

on the dielectric bombarded. This point will attain the chosen 
screen potential. 

A third system involves change of cathode potential with the 
screen potential held fast. The stable configuration is the 
curve of Fig. 14-9 with the digits and 1 recorded at the two 
stable points, where the dielectric assumes the cathode and screen 
potentials. To charge a spot originally at either of these stable 
potentials to the higher stable potential, the cathode potential is 
decreased to a value so low that its critical potential falls below 
the cathode potential in the stable condition, and then the elec- 
tron beam is aimed at the point to be charged. Under these 
conditions the potential of the point to be charged is above the 
critical potential for this electron beam, and this point takes the 
potential equal to the screen potential. After sufficient time, 
the cathode potential may be returned to its stable value, and 
the potential of the point being bombarded remains at the screen 
potential. To charge a spot originally at either of the stable 
potentials to the lower stable potential, the cathode potential 
is increased to a point below the stable critical potential but so 
high that the secondary-emission curve takes the form of Fig. 
14-10, rather than that of Fig. 14-9. The cathode ray is then 
aimed at the point in question. The potential of this point will 
drop to that of the cathode, and if the cathode is gradually 
returned to its original stable potential, the potential of the 
points of the dielectric will follow. 

In the description above, the principal idealization, which is 
not entirely valid, lies in assuming that each region of the dielec- 
tric can be charged independently of its neighboring regions. 
The description of the stable potentials under the curves shown 
would be valid for a target if the whole target were held at a 
single potential; however, if there are two potentials stored at 
different places on the target, it is clear that an infinite gradient 
cannot exist between them and that leakage in the target will 
produce a boundary region with potentials varying continuously 
from one of the stable potentials to the other. The existence of 
this boundary region was ignored in the discussion, but its 
behavior is of extreme importance. There is further influence 
of one region upon another by virtue of the small field generated 
by the charges at the various regions. 



362 PHYSICAL COMPONENTS AND METHODS 

Reading Principles. Several reading principles have been 
suggested for use in electronic static storage tubes. Of these, 
the most popular seems to be reading by means of the capture of 
secondary-emission electrons. This can be brought about in 
several ways, particularly for binary storage. The means 
chosen for any particular tube depends basically, of course, upon 
the choice of the method of using the stable points on the second- 
ary-emission curves for storage. 

For the first described method of utilizing the secondary- 
emission characteristics, the screen and cathode potentials are 
held constant at all times. The two stable potentials of the 
dielectric, the cathode potential and the screen potential, are 
used to store the binary digits and 1. An additional capacitor 
plate is provided, and a charge introduced on it will change the 
potential of the dielectric. If one point of the dielectric is 
irradiated and is at a stable potential, the number of electrons 
collected is exactly equal to the number of electrons furnished by 
the electron gun. On the other hand, if the potential of the 
target is decreased slightly by means of a charge introduced 
on the condenser plate, and if the stable potential of the dielectric 
before this charge was introduced was the screen potential, the 
displacement in potential will increase the secondary-emission 
ratio of the dielectric, as shown in Fig. 14-9. Momentarily the 
number of electrons collected will increase. Thus, the flow of 
current to the collector element will increase under such a dis- 
placement of potential if the stable potential was at its higher 
value. For the lower stable potential, however, this displace- 
ment of potential will have a negligible effect on the secondary- 
emission ratio, as shown in Fig. 14-9. Thus, displacement in 
this fashion gives rise to two signals which differ depending upon 
the original potential of the target ; this gives an adequate reading 
system, and fairly good signal-to-noise ratio has been obtained. 

For the second storage scheme described, where only the higher 
stable potential is used and the value of this potential is varied by 
changing the screen potential, a somewhat different reading 
system may be used. If only two values are stored, one cor- 
responding to screen potential E , and one corresponding to 
screen potential E h a reading may be made by setting the screen 
to a potential between these two values. If this is done, the 
characteristics of Fig. 14-9 are still valid. At the time of reading, 



TRANSFER MEDIUMS 363 

if the higher potential is stored, the target potential is at a point 
above the stable point on the curve, and the secondary-emission 
ratio is less than 1. On the other hand, if the stored potential 
is the lower stable potential, the secondary-emission ratio is 
greater than 1. It is to be noted that this reading scheme 
destroys the information stored, whereas the scheme described 
earlier leaves the stored information intact. 

For the third system of storage described above, in which the 
cathode potential is varied in order to bring about storage at the 
two stable potentials shown in Fig. 14-9, reading can also be 
brought about by variation of the cathode potential. The 
potential of the cathode is lowered to such an extent that the 
previous cathode potential {i.e., the lower stable storage poten- 
tial) falls to approximately the position of the minimum value 
of the secondary-emission ratio on Fig. 14-9. The screen poten- 
tial remains unchanged. When this is done, the secondary- 
emission ratio is decreased almost to zero for the lower stable 
storage potential, and it remains at approximately 1 for the 
higher stable storage potential. This permits reading without 
destruction of the stored information. 

Other schemes for reading stored information have been sug- 
gested. An example is an ingenious scheme tried by Rajchman 
in connection with the Selectron, which is described in Sec. 
14-2-9. He used anodized aluminum for the dielectric, with a 
hole drilled in the aluminum at the center of each of the regions 
into which the grids divided it. An anode element was placed 
behind this dielectric screen, and the reading principle was based 
on the grid action of the dielectric screen itself. With the beam 
directed at a region of the screen, the screen functioned as a 
blocking grid when at its lower stable potential in this region and 
as a passing grid at its higher stable potential. This scheme was 
found to operate satisfactorily, with suitable choice of param- 
eters, but the use of discrete metallic elements has led to simpler 
and more reliable operation, as will be explained later (see page 
371). 

A number of workers have suggested that continuous radiation 
on the dielectric be provided to hold it at a stable potential, 
thereby overcoming leakage difficulties. Where stability is 
attained in this fashion, patterns have been stored for many 
hours without noticeable deterioration. 



364 PHYSICAL COMPONENTS AND METHODS 

MIT Storage Tube. A number of electrostatic storage tubes 
having storage capacities from 256 to 1,024 digits have been 
constructed and used successfully at the MIT Servomechanisms 
Laboratory. A bank of such tubes will be used as the high-speed 
storage system in the Whirlwind I computer. 

The MIT tube 60-62 has two cathode-ray guns, a high-velocity 
gun for use in recording and reading data, and a low-velocity 
gun to supply a diffused electron flood for the purpose of main- 
taining the charge pattern, or array of stored data. 

The storage surface is a mosaic of conducting beryllium, 
sputtered on a dielectric surface. There is a collector screen 
between the guns and the storage surface. A signal plate 
beyond the beryllium storage surface is insulated from it by the 
dielectric sheet. 

The potentials of the guns and the screen are not varied. The 
signal plate is switched between screen potential (ground) and a 
positive potential of about 100 volts. 

Writing Principle. The two stable conditions corresponding 
to the digits and 1 are zero charge and negative charge on the 
storage surface. The recording of either one or the other of these 
stable conditions upon a particular area on the surface is accom- 
plished by pulsing the high-velocity gun with the signal plate at 
either ground potential or at positive potential. Regardless of 
the potential of the signal plate, the irradiated area on the 
storage surface will assume the potential of the collector screen 
(ground). However, if the signal plate is at positive potential 
at the instant of irradiation, then, when the signal plate potential 
is returned to ground, the irradiated area becomes negative 
(representing the digit 1). 

Holding. During the time when the high-velocity gun is not 
being pulsed, the low-velocity holding gun produces a continuous 
stream of electrons. If a spot of zero potential on the storage 
surface (corresponding to the digit 0) becomes slightly positive, 
it will collect electrons from the holding-gun beam until it 
returns to collector potential. If the same spot becomes slightly 
negative, electrons will be lost from the surface until it again 
reaches the collector potential. Hence, the spot at zero potential 
is in a stable condition. 

A negative spot (corresponding to the digit 1) is stable also. 
Since it tends to be at the same potential as the holding-gun 



TRANSFER MEDIUMS 365 

cathode, electrons will ordinarily be decelerated as they approach 
and will arrive with low or zero velocity. If a negative spot 
becomes slightly more positive than the holding-gun cathode, the 
storage surface will collect electrons faster than they are being 
emitted from it and will become more negative. 

On the other hand, if the negative spot becomes more negative 
than the holding-gun cathode, primary electrons will be reflected. 
Because of leakage, the negative spot will gradually become more 
positive. 

Reading. Reading is accomplished by examining the change 
in the voltage of the signal plate when a particular spot on the 
storage surface is irradiated by the high-velocity gun. Irradia- 
tion of a spot at ground potential (representing a 0) produces no 
change and no output. When a negative spot is irradiated, 
secondary electrons are given off faster than primary electrons 
are striking it; the spot becomes more positive. This effect 
produces a positive output signal on the signal plate which is 
capacitively coupled to the storage surface. 

The NRL Tube. The storage tube developed by Dr. Haeff 
at Naval Research Laboratory 63 ' 64 resembles a conventional 
cathode-ray tube and employs the screen-voltage writing tech- 
nique described above. To place a negative charge upon the 
target, a large positive pulse is applied to a control grid located 
near the electron gun at the same time a negative pulse is applied 
to the collector screen. The positive voltage on the control grid 
increases the intensity of the beam, and the negative voltage on 
the screen decreases the velocity of the stream. Thus, an 
intense beam of slow electrons deposits a negative charge on 
the dielectric target, reducing its potential to that of the cathode. 
In order to place a positive charge upon the target, positive pulses 
are simultaneously applied to both control grid and collector 
screen. A high-intensity, high-velocity electron beam results 
and causes secondary emission from the point on the target being 
bombarded. The net flow of electrons away from the target 
brings its potential up to that of the screen, and it tends to stay 
there after the pulses have been removed. 

A diffused holding beam is irradiated over the target to main- 
tain the established charge pattern. The potential of the hold- 
ing-beam gun is such that positively charged spots are above the 
critical potential and negatively charged spots are below the 



366 PHYSICAL COMPONENTS AND METHODS 

critical potential. The effect of the electron spray from the 
holding gun is to hold the negative spots at cathode potential 
and the positive spots at screen potential. In order to keep the 
velocity of this beam nearly constant, the pulses applied to the 
collector screen are also applied to the gun from which the 
holding beam emanates. Thus, the relative potential remains 
constant. 

In a present model of this tube, the same gun is used for reading 
and writing. In order to minimize noise caused by modulating 
pulses and switching operations during the reading process, the 
reading beam is modulated at an intermediate frequency. Since 
the output signal from the collector screen is modulated, it can 
be passed through a tuned amplifier which rejects the low fre- 
quencies, including the noise. The output of the intermediate- 
frequency amplifier is rectified and introduced to a video ampli- 
fier where l's and O's appear as positive pulses of different 
amplitudes. These may be balanced with negative pulses of 
such an amplitude that a negative pulse will be received for a 
negative spot on the target, and a positive pulse will be received 
when a positive spot is probed. 

University of Manchester Tube. Another deflection-type 
electrostatic storage tube is being developed at the University 
of Manchester, England, and has been reported by Williams and 
Kilburn. 65 Their work has been done on commercial cathode- 
ray tubes, which are modified by the addition to the large end of 
the envelope of a metallic plate which serves as a signal plate. 
The glass and the phosphorescent screen coating become the 
dielectric target. In the tube described, the deflection plates, 
the internal conductive coating, and the first and third anodes 
are all at ground potential; the cathode, the grid, and the focus 
electrode are given a negative potential in the order of —2,000 
volts. Before an electron beam is initiated, the inside of the 
target tends to assume ground potential because of leakage to 
the grounded coatings on the inside of the envelope. Since 
there is no collector screen, drifting electrons are caught by the 
anodes or the deflecting structure, and the secondary-emission 
characteristics are similar to those illustrated in Fig. 14-8; 
the cathode potential is near the middle of the range in which 
the secondary-emission ratio is greater than 1. Thus when a 
beam is initiated and aimed at a spot on the screen, there is a net 



TRANSFER MEDIUMS 



367 



flow of electrons away from the spot for a short period of time 
until equilibrium conditions are reached again. 

In order to understand the method employed for representing 
binary digits on the target of this tube, it is necessary to consider 
the charge distribution and the charge-time history in more 
detail than was required for the tubes previously described. As 
electrons leave the spot bombarded, the spot becomes positive 
in potential so that the secondary electrons emitted from the 
target are attracted back to it by a field which is increasing for a 




Fig. 14-11. An electrostatic well. The curve illustrates voltage distribu- 
tion along a line in the plane of the tube face. The downward direction is 
positive. 

short time after the primary beam is initiated. Thus, more and 
more electrons return to the target, and equilibrium is reached 
when the number escaping the target is equal to the number in 
the primary stream. This is the condition of unity secondary- 
emission ratio, described earlier. The electrons which leave the 
target with higher velocities have time to acquire a velocity 
parallel to the face of the target so that some of those which 
return lie around the rim of the original spot area. Thus the 
potential along a line through the center of the spot which 
has been bombarded is illustrated by Fig. 14-11. Such a charge 
distribution has been called a well. Experiments have shown 
that if the bombardment does not continue for more than 400 
microseconds, the target is not affected at distances greater 
than a spot diameter from the center of the spot. 



3G8 



PHYSICAL COMPONENTS AND METHODS 



When a spot on a target near the rim of an existing well is 
bombarded, there is a tendency for secondary electrons to fill the 
original well. The original well is not filled as fast as the second 
well is formed, however; hence after a short period of time the 
potential along a line through the centers of the holes is as shown 
in Fig. 14-12. 

Two spots placed side by side on a target will be mutually inde- 
pendent if their spacing is more than a critical value, about 1.33 
diameters, but they will affect each other if their spacing is less 
than this critical distance. Use is made of this phenomenon in 
order to make two kinds of distinguishing marks on the target. 




Fig. 14-12. Double electrostatic well. 



One mark is a full well, which is made by bombarding a single 
spot for a short time. The other mark is a well which is modified 
by bombarding another spot in the near vicinity of the storage 
spot. These marks will be maintained for a few tenths of a 
second, depending on surface leakage. Therefore, it is necessary 
to read and rewrite the marks continually. Several methods 
have been tested for reading and differentiating between marks 
of this type. Reading is based on the fact that when an electron 
beam is aimed at a full well and initiated, the presence of the 
negative cloud of electrons in the primary beam and in the 
secondary-electron stream induces an instantaneous flow of 
current to the metallic plate on the outside of the tube. This 
plate is connected to the input of an amplifier. An output wave 
form is obtained which is dependent upon the transient response 
of the amplifier to the negative pulse received at the input. 



TRANSFER MEDIUMS 369 

When a beam is turned on a storage spot where the well has been 
modified by the presence of a second well nearby, there is a 
redistribution of charge on the face of the tube, which induces a 
current in the input circuit of the amplifier, this current being 
of opposite polarity to that of the current induced by the electron 
cloud. If parameters are adjusted properly, the pulse of positive 
polarity induced by the shifting charge may be made large enough 
to counterbalance the negative pulse induced by the electron 
cloud. Thus the type of mark stored determines the polarity 
of pulse received at the plate when the spot is bombarded by an 
electron beam. 

A regenerative system using a gate and a pulse generator can be 
used to renew each mark periodically after it is read. It is 
proposed to scan this tube in a manner similar to that used with a 
television tube, so that regeneration can be accomplished. In 
addition, a modification of this sequence can be made in order to 
read or write a digit or line of digits immediately after orders to 
do so are generated by the computer. If scanning is from left 
to right, the left-hand is modified by the right-hand well when 
the double mark is written. When the mark is regenerated, the 
right-hand well is modified by the left-hand one momentarily, 
but finally as the beam of electrons moves to the second well, 
the left-hand well is again modified by the right-hand well. 

A space approximately two by three spot diameters (d) on the 
target is required to store each mark, so if d = 1 millimeter, a 
6-inch tube will store about 600 digits. Improving the focus of 
the beam increases the storage capacity as the inverse square of 
the spot diameter. 

Other Electrostatic Storage Devices. Zworykin and Morton 67 
have described a type of storage tube in which one electron 
beam writes a potential-variation pattern on. one side of a two- 
sided mosaic while another electron beam scans the other side 
to release electrons, which produce a variable-current signal in a 
collector anode. Such a tube might be modified to include an 
electron multiplier for amplifying the collected current. A 
storage tube called the Krawinkel 66 was devised in Germany 
during the war. Apparently it uses an Iconoscope type of 
mosaic for the storage medium. 

That the ultimate storage potentialities of such devices may be 
far in excess of 4,096 digits per tube is suggested by the present 



370 PHYSICAL COMPONENTS AND METHODS 

commercial-television picture standards. The RMA standard 
picture contains some 360,000 elements (525 lines, 3 by 4 aspect 
ratio), and satisfactory experimental pictures have had over a 
million elements. Although the elements are by no means com- 
pletely resolved, these figures indicate what may be expected in 
future refinements of electronic image and storage devices. 

14-2-9. The Selectron. An electrostatic storage tube called 
the Selectron, the latest model of which has been designated 
SE256, is at the time of this writing in an advanced stage of 
development at the RCA Laboratories in Princeton, N.J. 63-65 
This tube has a storage capacity of 256 binary digits. It is 
3 inches in diameter, 7 inches long, and utilizes a 40-lead stem. 
Access time to any element is 20 microseconds and is independent 
of all previous accesses to other elements. 

A number of these tubes have been made already for experi- 
mental use; uniformity of selection and control characteristics 
have been observed in all the tubes. 

Principle of Operation. The Selectron utilizes the fact that an 
insulated secondary-electron emitter can be made to "float" at 
either of two stable potentials under certain conditions of 
primary-electron bombardment and secondary-electron collec- 
tion. Secondary electrons are received by a collector electrode 
placed near the insulated secondary emitter in a manner similar 
to the collector screen of the deflection-type storage tubes. 
While a surface element is being irradiated, its potential is locked 
at one of these stable values. 

As originally conceived, the Selectron consisted essentially of a 
cylindrical dielectric target with the cathode at its axis and the 
electron stream flowing radially from the cathode. There was a 
grid, made of 64 vertical metallic bars and 65 circular horizontal 
metallic rings, just inside the target. The 64 2 , or 4,096, windows 
in this cylindrical grid divided the target into 4,096 discrete 
cells. The size of the windows was such that electrons could 
pass through a window only when the four metallic conductors 
bounding that window were all more positive than some particu- 
lar voltage. 

The physical form of the tube has been altered materially 
since it was first conceived. As a result of Rajchman's experi- 
mental work, directed primarily toward improvements in relia- 
bility, he has adopted a flat target (in place of the original 



TRANSFER MEDIUMS 371 

cylindrical form) and eight elongated cathodes. In the SE256 
there are two parallel target planes on opposite sides of the plane 
of the cathodes. A detailed description of the engineering 
features of this tube has been published. 65 

Regardless of the physical form, the operation of the Selectron 
depends upon the effects obtained by energizing the four con- 
ductors bounding a particular window. One of its greatest 
advantages is derived from the fact that the means for directing 
electrons into a selected cell does not depend upon the accurate 
reproduction of deflection voltages. 

In the tube's quiescent operating state all the selecting bars 
are at the potential of the cathodes (zero). Electrons emitted 
from the cathodes are focused into 256 beams by the combined 
action of these bars (at zero potential) and a collector plate which 
is held at some positive potential such as 180 volts. These beams 
are focused on 256 nickel-plated steel eyelets, which are floating 
electrically and are fixed in mica sheets. There are only two 
stable voltage conditions at which an eyelet can be held, because 
of the secondary-emission phenomena. Every one of the 256 
eyelets will remain in one or the other of these states so long as the 
tube is in its quiescent operating state. 

To write into a particular element, current is interrupted every- 
where except to it. Then a signal is applied to a perforated 
metal plate, called the writing plate, which is located (with respect 
to the cathodes) beyond the mica plate in which the eyelets are 
set. This plate is capacitively coupled to all the eyelets. The 
voltage on the particular eyelet will be left at either one or the 
other of its two stable levels depending upon the presence or 
absence of an additional pulse on the selecting bars. This pulse 
cuts off the current to the selected eyelet during the decay of a 
pulse applied to the writing plate. (This pulse is positive for 
either polarity writing.) The eyelet will be left at the higher 
level if only a single positive pulse is applied and at the lower 
level if the initial positive pulse is followed by an additional pulse 
to one or more of the four bars bounding its particular window. 

Reading is effected by obtaining an indication as to whether 
current is passing through the central hole in the eyelet. When 
an eyelet is positive, some of the electrons directed at it go 
through the hole. When it is negative, electrons are repelled and 
do not go through it. An auxiliary electrode, called the reading 



372 PHYSICAL COMPONENTS AND METHODS 

plate, normally biased negatively, cuts the potentially large cur- 
rent from all or a large proportion of the eyelets to the output 
electrode. To read, an element is selected, then the reading 
electrode is pulsed positively. There is, or there is not, a result- 
ing pulse of current in the output electrode, depending on the 
potential of the interrogated eyelet. The output electrode is a 
set of wires well shielded from capacity pickup by a Faraday 
cage, which also serves to support a fluorescent screen for moni- 
toring purposes. 

The Selectron Aiming Principle. It is of interest to note that 
the aiming method described is not the only one which might 
be applied to a system of bars and windows. Therefore, a more 
general discussion of the Selectron aiming principle is included 
here, as an appendix to the foregoing material. Consider a 
circular picket fence with 2 n pickets (n integral and at least 2) 
and 2 n spaces between pickets. Let these pickets be numbered 
in their cyclic order from to 2 n — 1 with a number i. Assign 
to the space following a picket the same number assigned to a 
picket. 

Now choose two positive integers p and q, such that 

p + q = n 

Assign to each even-numbered picket a number j (varying from 
picket to picket) with < j < 2 P — 1 ; to each odd-numbered 
picket, assign a number k with < k < 2 q — 1. It is clear 
that for any given space between two pickets there are 2 P ways 
the adjacent even-numbered picket can be labeled (because 
there are 2 P different numbers j); and 2 q ways the adjacent odd- 
numbered picket can be labeled (because there are 2 q different 
numbers k). Then, for any space there are 

2 p X 2 q = 2 p+q = 2 n 

ways the two adjacent pickets can be labeled. This is exactly 
the number of spaces available to be labeled. The Selectron 
principle of aiming is based on a labeling scheme in which every 
possible labeling of pairs of successive pickets occurs exactly 
once. 

One way in which this can be done is as follows: First, let us 
assign the numbers through 2 P — 1 cyclically to the pickets 



TRANSFER MEDIUMS 373 

with even numbers; i.e., assign numbers j to the pickets with 
even numbers i as shown in Table 14-4. 

Table 14-4 

* 3 



2 1 

4 2 



2 X 2p - 2 2p - 

2X2" 

2 X 2p + 2 1 



4 X 2? - 2 2* - 

4X2" 



6 X 2" - 2 2p - 1 

6X2" 



2 n - 2 = 2* X 2p - 2 2p - 1 

It is seen that for each j cycle the number i increases by 
2 X 2 P = 2 P+1 . Then, the number of j cycles must be 

_ Og-1 

2 P+1 ~ L 

To the odd-numbered pickets appearing in each of these cycles 
let us assign two values of k alternately, using a different pair of 
values in each j cycle. In the first j cycle the values k = and 
k = 1 are assigned alternately ; in the second the values k = 2 and 
k = 3, etc., until the (2«- 1 )th, which has the values k = 2 q — 2 
and k = 2 q — 1 assigned alternately. The whole assignment is 
given in Table 14-5, a being the j-cycle number, from to 

2<r-i - 1. 

It is clear that this scheme has the desired property. Indeed, 
any chosen value of k occurs in exactly one j cycle, and at some 
point in this cycle it is adjacent to each j value. Thus, given 



374 



PHYSICAL COMPONENTS AND METHODS 



any admissible j value and any admissible k value, there is one space 
with these labels on its pickets. 



Table 14-5 



a 


i 


j 


k 

















1 


— 








2 


1 


— 





3 


— 


1 





4 


2 


— 





5 


— 








6 


3 


— 





7 




1 





2 X 2^ - 2 


2" - 1 







2 X 2* - 1 


— 


1 




2X2" 





— 




2 X 2*> + 1 


— 


2 




2 X*2" + 2 


1 


— 




2 X 2" + 3 


— 


3 




2 X 2" + 4 


2 


— 




2 X 2" + 5 




2 


2«-i _ i 


2 n - 4 


2^-2 




2«-i - 1 


2" - 3 


— 


2i - 2 


29-i _ i 


2* - 2 


2" - 1 


— 


29-i _ i 


2" - 1 


— 


2i - 1 



14-3. Transfer in Space 

It is evident that data must be transferred from one unit in a 
computer to another while a problem is being solved. In elec- 
tronic digital computers, data are transformed into electrical 
pulses which are transferred along busses and cables and are 
initiated by the action of buffers and gates. All of these four 
components are discussed in previous sections and are con- 
sidered here only to the extent that the problem of transferring 
data in space affects the requirements which must be met in 
their design. 



TRANSFER MEDIUMS 375 

14-3-1. General Requirements. The operational require- 
ments for transferring pulses in a computer are: 

1. Freedom from transfer failure due to: 

a. Undesired inductively or capacitively coupled signals. 

b. Cross modulation. 

c. Spurious noise. 

d. Variations in components caused by aging, atmospheric 
• conditions, or from any other predictable cause. 

2. Limited phase, frequency, and amplitude distortion. 

3. A minimum of time delay. 

4. Limited reflection of pulses caused by discontinuities and 
mismatches. 

In addition, the devices employed should be economical and 
easy to service. 

The designers of the ENIAC have taken adequate precautions 
to observe requirement 1. The average signal level is between 
25 and 50 volts. A normal variation in component characteris- 
tics is not harmful. To guard against high cross-coupling volt- 
ages, the impedance levels of the transmission lines have been 
kept low, and connections between units of the machine have 
been made through shielded microphone cables. The precau- 
tions to be observed in tube and wiring layout inside units are 
similar to those of video amplifier design. 67 ' 71 * 72 ' 75,76 Compo- 
nents should be arranged to minimize lead length, and critical 
leads should be spaced apart or transposed frequently. Imped- 
ance levels should be kept as low as possible. In the ENIAC, 
circuits have been designed so that wide variations in parameters, 
and particularly vacuum-tube variations, do not interfere with 
the positive, dependable operation of the unit. 

Requirements 2 and 3 are familiar to designers of radio-fre- 
quency amplifiers and transmission networks and are, essentially, 
the problems discussed in Sec. 14-2-7 regarding time delay and 
distortion in transmission lines. Distortion in the transfer 
circuits must be limited sufficiently so that any combination of 
pulses is perfectly distinguishable at the receiving unit. Time 
delay, on the other hand, is somewhat less objectionable than 
distortion since it can be compensated by adding equivalent 
delays in other circuits in the unit. However, in the interest of 
simplicity, it is well to eliminate the necessity for corrective- 



376 



PHYSICAL COMPONENTS AND METHODS 



delay circuits wherever possible. In circuits which require a 
high impedance level, shunt capacity should be minimized, and 
if circuits are frequency-selective within the frequency range of 
interest, frequency compensation may be required. A familiar 
example is that of inductive compensation in the plate load of a 
vacuum-tube amplifier, as illustrated in Fig. 14-13. The high- 
frequency components of a pulse are partially shunted by 
capacitor C, which represents the effective sum of all the capaci- 
ties in parallel with the plate resistor. A properly chosen* value 



INPUT 




COMPENSATED 



UN-COMPENSATED 



'OUT ""^strat" 

.42 R 2 C 

DAMPING RESISTANCE TO 



Fig. 14-13. 



B+ PREVENT OVERSHOOT 

Inductive compensation of plate capacitance. 



of inductance L will compensate for some of this shunting by 
holding the plate load impedance to a constant value for a greater 
portion of the frequency range. More sophisticated forms 
of phase and frequency compensation are discussed in the 
literature. 71 

Another source of frequency discrimination is in the interstage 
coupling capacitors, which affect the low-frequency response of 
the units. To overcome this difficulty, many of the circuits in 
the ENIAC are direct-coupled. Direct coupling introduces 
another difficulty, however. In the ENIAC, largely because of 
the direct coupling between stages, there are 78 different d-c 
voltages which must be held constant within a few per cent. In 
some video amplifiers, d-c restoring circuits are employed to 
avoid the selective characteristics of a-c coupling without the 
necessary complications of d-c coupling. Undoubtedly these 
circuits will find use in future high-speed computers. 

In the ENIAC it was not necessary to terminate transmission 
cables in their characteristic impedances in order to reduce 
reflections, but as computers using shorter pulses are designed, 



TRANSFER MEDIUMS 



377 



it becomes necessary to satisfy this requirement. Moreover, at 
the time of this writing, it has been found desirable to terminate 
the lines in the ENIAC in impedance-matching load boxes for 
the sake of greater reliability, although the machine will operate 
without them. 

14-3-2. Gates and Buffers. Gates and buffers have been 
discussed in Chap. 4, and as such the multicontrol grid, the 
parallel load, and the crystal gates were described. 73 In applica- 




>OUT 



1 -(G9f 



Fig. 14-14. Multiple-gate circuit. 

tions where it is necessary to use gating or buffer circuits having 
a multiplicity of inputs, the accumulated shunt capacity at the 
plate circuits of a number of triodes connected in parallel limits 
the leading-edge rise time of pulses to be passed. A circuit 
known as the christmas-tree circuit and illustrated in Fig. 14-14 
is useful in reducing this output capacity. A block diagram of 
the circuit resembles a tournament play-off diagram. For 
example, circuit a consists of two-input gates for a minimum of 
capacity loading. A larger number of elements can be used in 
parallel at the expense of circuit response and/or gain. A 
similar circuit such as Fig. 14-15 might also be devised using 
multiple-grid tubes. 

A miniature pentode tube which is suitable for use as a gate is 
the 6AS6 manufactured by Western Electric. The control grid 



378 PHYSICAL COMPONENTS AND METHODS 

and the suppressor grid are used as the independent control 
elements. Some combinations of electrode voltages yield a 
suppressor-grid transconductance which is almost half that of 
the control grid and nearly equal cutoff voltages for the two 
grids. In most pentodes, the control-grid transconductance 
is many times the suppressor-grid transconductance. It does 
not seem feasible to increase the number of control grids beyond 
two, because in order to make grids of nearly equal sensitivity, 
it is necessary to decrease the sensitivity of all grids to that of 
the one which is spaced farthest from the cathode. The result 
is a tube of very low transconductance for each control grid. 



^& 



IN 



i^& 



?:4H- 



t^&n 



G6 



G7 ► OUT 




Fig. 14-15. Multiple-gate circuit. 

A study has been made at Harvard 74 of multiple-grid tubes for 
use as multiple-input gates. A fundamental difference in require- 
ments has been established between gates which are to handle 
large voltage pulses and those which receive small voltage 
pulses. For the former case, the grid is driven from anode cutoff 
to anode saturation by the pulses. Sharp cutoff with high 
average transconductance is desired in order that output pulses 
of steep wave front may be obtained. In multiple gates which 
receive small input pulses, the ratio between on and off trans- 
conductance should be large. When a large number of gate 
tubes are used with a common plate load resistor, there is a 
tendency for the spurious noise signals on grids biased to cutoff 
to add up to an amount comparable to an output pulse. For 
this reason a high ratio between on and off transconductance is 
desired. A diode gate is shown in Fig. 14-16 which is a slight 
variation of the one illustrated in Fig. 4-26 and described on 
page 41. Resistor Ri is much greater than R 2 , with the result 
that input pulses are effectively shorted through one diode or 



TRANSFER MEDIUMS 379 

the other, depending upon pulse polarity. Application to R 2 
of a gate voltage of the polarity indicated places a bias on the 
two diodes so that input pulses are not shunted and appear at 
the output. Actually a small voltage gets through this type of 
gate whether it is opened or not, and the ratio of voltage in the 
on and off conditions is limited to the ratio of backward to 
forward resistance through the crystal rectifiers. 

High-speed Coincidence Devices. A variation of the Rossi 
circuit described in Sec. 4-3-3 and illustrated in Fig. 4-16 on 



IN r-4 r-4 OUT 




Fig. 14-16. Diode gate. 

page 38 has been developed in connection with nuclear experi- 
mentation. 77 This reference is noteworthy in this text because 
it describes by far the fastest use of a coincidence or and circuit 
known to the authors. The resolving power reported is of the 
order of 10 -8 to 10 -9 seconds. In this device a tube equivalent 
to the Type 6AC7, having high mutual conductance and low 
capacity, is used. 

A method of detecting coincidences of very short duration and 
observing them on a cathode-ray screen also has been devised 
and reported. 79 

REFERENCES 

Relay Storage 

1. M.I.T. Radiation Laboratory Staff, Relay Data Including Shock and 
Vibration Measurements, Radiation Laboratory Report 747 (Aug. 1, 1945). 

2. Rosslyn, J., Electric Relays (Chemical Publishing Company, Inc., 
Brooklyn, 1921). 

Punched-tape Storage 

3. Alexander, Samuel N., Input and Output Devices for Electronic Digital 
Calculating Machinery, Annals of the Computation Laboratory of Harvard 
University, Vol. XVI, pp. 248-253 (1948). 



380 PHYSICAL COMPONENTS AND METHODS 

4. Lake, Ross A., Teletype Reperforator Transmitter for Automatic 
Switching Systems, Bell Laboratories Record, Vol. 26, No. 3, pp. 106-110 
(March, 1948). 

5. Miller, Frederick G., Application of Printing Telegraph Equipment to 
Large-Scale Calculating Machinery, Annals of the Computation Labora- 
tory of Harvard University, Vol. XVI, pp. 213-222 (1948). 

6. Sparks, S., and R. G. Kreer, Tape Relay System for Radio-Telegraph 
Operation, RCA Review, Vol. 8, No. 3, pp. 393-426 (September, 1947). 

7. Williams, S. B., Bell Telephone Laboratories Relay Computing System, 
Annals of the Computation Laboratory of Harvard University, Vol. XVI, 
pp. 41-68 (1948). 

Photographic Storage 

8. Carlson, F. E., and D. A. Prichard, Flash Tubes, Illumination Engineer- 
ing, Vol. 42, pp. 235-248 (February, 1947). 

9. Mees, C. E. Kenneth, The Theory of the Photographic Process (The 
Macmillan Company, New York, 1942), pp. 838. 

10. O'Neal, R. D., Photographic Methods for Handling Input and Output 
Data, Annals of the Computation Laboratory of Harvard University, Vol. 
XVI, pp. 260-266 (1948). 

11. O'Neal, R. D., and A. W. Tyler, Photographic Digital Reader- Recorder, 
Progress Report 1 (Eastman Kodak Co., Rochester, N.Y., 1948). 

12. Tyler, A. W., Optical and Photographic Storage Techniques, Annals of 
the Computation Laboratory of Harvard University, Vol. XVI, pp. 146-150 
(1948). 

Magnetic Storage 

13. Aldous, D. W., A Bibliography of Magnetic Recording, Electronic 
Engineering, Vol. 19, pp. 390-391 (December, 1947). 

14. Barrett, A. E., and C. J. F. Tweed, Some Aspects of Magnetic Record- 
ing, Journal of the Institution of Electrical Engineers (London), Vol. 82, 
No. 495, pp. 265-285 (March, 1938). 

15* Begun, S. J., Recent Developments in Magnetic Sound Recording, 
Journal of the Society of Motion Picture Engineers, Vol. 48, pp. 1-13 
(1947). 

16. Begun, S. J., Magnetic Recording and Some of Its Applications in 
the Broadcast Field, Proceedings of the Institute of Radio Engineers, 
Vol. 29, No. 8, pp. 423-433 (August, 1941). 

17. Begun, S. J., Magnetic Recording (Murray Hill Books, Inc., New York, 
1949). 

18. Boenning, W. R., Magnetic Recording and Reproduction of Pulses (Engi- 
neering Research Associates, Inc., Arlington, Va., 1947). 

19. Bozorth, R. M., Magnetism, Reviews of Modern Physics, Vol. 29, No. 1, 
pp. 29-86 (January, 1947). 

20. Camras, M., Recent Developments in Magnetic Recording for Motion 
Picture Film, Journal of the Acoustical Society of America, Vol. 19, pp. 
322-325 (March, 1947). 

21. Camras, M., Theoretical Response from a Magnetic-Wire Record, 



TRANSFER MEDIUMS 381 

Proceedings of the Institute of Radio Engineers, Vol. 34, pp. 597-602 
(August, 1946). 

22. Cohen, A. A., and W. R. Keye, Selective Alteration of Digital Data in a 
Magnetic Drum Computer Memory, unpublished (P^ngineering Research 
Associates, Inc., St. Paul, Minn., 1948.) Pamphlet. 

23. Coombs, J. M., Storage of Numbers on Magnetic Tape (Engineering 
Research Associates, Inc., St. Paul, Minn., 1947). Pamphlet. 

24. Hansel, C. W., et at., Report on the Magnetophone, OTS Report PB 1346 
(1945). 

25. Hickman, C. N., Sound Recording on Magnetic Tape, Bell System 
Technical Journal, Vol. 16, pp. 165-177 (1937). 

26. Holmes, L. C, and D. L. Clark, Supersonic Bias for Magnetic Recording, 
Electronics, Vol. 18, pp. 126-136 (July, 1945). 

27. Korhone, E. D., L. W. Reid, and W. L. Anderson, Preparation and 
Characteristics of Magnetic Recording Surfaces, pamphlet prepared by 
Engineering Research Associates, Inc., under Bureau of Ships contract 
NObsr 42001 (Jan. 7, 1949). 

28. Kornei, Otto, Survey of Magnetic Storage, Annals of the Computation 
Laboratory of Harvard University, Vol. XVI, pp. 223-237 (1948). 

29. Long, T. H., and G. D. McMullen, B-H Curve Tracer for Magnetic 
Recording Wire, Transactions of the American Institute of Electrical 
Engineers, Vol. 65, pp. 146-149 (1946). 

30. Long, T. H., A New Wire Recorder Head Design, Transactions of the 
American Institute of Electrical Engineers, Vol. 65, pp. 216-219 (1946). 

31. Massachusetts Institute of Technology, Staff of the Department of 
Electrical Engineering, Magnetic Circuits and Transformers (John Wiley 
& Sons, Inc., New York, 1943), pp. 3-40. 

32. Menard, J. Z., High Frequency Magnetophone Magnetic Sound Recorders, 
OTS Report PB 12659 (1946). 

33. Moore, B. L., Magnetic and Phosphor Coated Discs in Proceedings of 
a Symposium on Large-Scale Digital Calculating Machinery (Harvard 
University Press, Cambridge, Mass., 1948), pp. 130-132. 

34. Sears, A. W., Wire Recorder Wow, Journal of the Acoustical Society of 
America, Vol. 19, No. 1, pp. 172-178 (January, 1947). 

35. Selby, M. C, Investigation of Magnetic Tape Recorders, Electronics, 
Vol. 17, pp. 133-135 (May, 1944). 

36. Shaper, H. B., FM Magnetic Tape Transient Recorder, Proceedings 
of the Institute of Radio Engineers, Vol. 33, pp. 753-760 (November, 
1945). 

37. Sheppard, C. B., Transfer between External and Internal Memory 
Annals of the Computation Laboratory of Harvard University, Vol. XVI, 
pp. 267-273 (1948). 

38. Toomim, H., and D. Wildfeuer, Mechanism of Supersonic Frequencies 
as Applied to Magnetic Recording, Proceedings of the Institute of Radio 
Engineers, Vol. 32, pp. 664-668 (November, 1944). 

39. Wooldridge, D. E., Signal and Noise Levels in Magnetic Tape Record- 
ing, Transactions of the American Institute of Electrical Engineers, Vol. 
65, pp. 343-352 (1946). 



382 PHYSICAL COMPONENTS AND METHODS 

Phosphor Storage 

40. du Pont, E. I., de Nemours & Co., Inc., Bibliography — Phosphorescence 
and Phosphorescent Sulfides, special report prepared in connection with 
research under Contract No. W3G-039-SC-32256 (Feb. 16, 1948). 

41. Moore, Benjamin L., Magnetic and Phosphor Coated Discs, Annals of 
the Computation Laboratory of Harvard University, Vol. XVI, pp. 130- 
132 (1948). 

Sonic-delay-line Storage 

42. Arenberg, D. L., Ultrasonic Solid Delay Lines, Journal of the Acoustical 
Society of America, Vol. 20, pp. 1-26 (January, 1948). 

43. Hunting, H. B., A. G. Emslies, and V. W. Hughes, Ultrasonic Delay 
Lines, Parts I and II, Journal of the Franklin Institute, Vol. 245, pp. 1-24, 
101-115 (January and February, 1948). 

44. Knott, G. C, Reflexion and Refraction of Elastic Waves with Seis- 
mological Application, Philosophical Magazine, Vol. 48, Series 5, pp. 
64-97 (July, 1899). 

45. Macelwane and Sohon, Introduction to Theoretical Seismology (John 
Wiley & Sons, Inc., New York, 1932, 1936). 2 vols. 

46. Metz, F. A., Jr., and W. M. A. Andersen, Improved Ultrasonic Delay 
Lines, Electronics, Vol. 22, No. 7, pp. 96-100 (July, 1949). 

47. Ridenour, L. N., Radar System Engineering (McGraw-Hill Book Com- 
pany, Inc., New York, 1947), pp. 667-677. 

48. Ringo, G. R., J. W. Fitzgerald, and B. G. Hurdle, Propagation of 
U.H.F. Sound in Mercury, Physical Review, Vol. 72, No. 1, pp. 87-88 
(July, 1947). 

49. Sharpless, T. Kite, Mercury Delay Lines as a Memory Unit, Annals of 
the Computation Laboratory of Harvard University, Vol. XVI, pp. 103-109 
(1948). 

50. Sharpless, T. Kite, Design of Mercury Delay Lines, Electronics, Vol. 20, 
No. 11, pp. 134-138 (November, 1947). 

Electromagnetic-delay-line Storage 

51. Blewett, J. P., and J. H. Rubel, Video Delay Lines, Proceedings of the 
Institute of Radio Engineers, Vol. 35, No. 12, pp. 1580-1584 (December, 
1947). 

52. Brillouin, Leon N., Electromagnetic Delay Lines, Annals of the Com- 
putation Laboratory of Harvard University, Vol. XVI, pp. 110-124 (1948). 

53. Kallmann, Heinz E., High-Impedance Cable, Proceedings of the Institute 
of Radio Engineers, Vol. 34, No. 6, pp. 348-351 (June, 1946). 

54. Kallmann, Heinz E., Equalized Delay Lines, Proceedings of the Institute 
of Radio Engineers, Vol. 34, No. 9, pp. 646-657 (September, 1946). 

55. Montgomery, Dicke, and Purcell, Principles of Microwave Circuits, 
Radiation Laboratory Series (McGraw-Hill Book Company, Inc., New 
York, 1948), pp. 203-206. 

56. Rudenberg, Rheinhold, Electromagnetic Waves in Transformer Coils 
Treated by Maxwell's Equations, Journal of Applied Physics, Vol. XII, 
pp. 219-229 (March, 1941). 



TRANSFER MEDIUMS 383 

57. Slater, J. C, Microwave Transmission (McGraw-Hill Book Company, 
Inc., New York, 1942). 

58. Trevor, J. B., Jr., Artificial Delay Line Design, Electronics, Vol. 18, 
No. 6, p. 135 (June, 1945). 

59. Zimmerman, K. H., Spiral Delay Lines, Electrical Communications, 
Vol. 23, No. 3, pp. 327-328 (September, 1946). 

Electrostatic Storage 

60. Forrester, J. W., Electrostatic Storage Tubes, MIT Servomechanisms 
Laboratory Report No. 6845-R-110 (Jan. 13, 1947). 

61. Forrester, J. W., Electrostatic Storage Tubes for Digital Computers and 
other Information-Processing Systems, MIT Servomechanisms Labora- 
tory Report No. R-153 (Jan. 26, 1949). 

62. Forrester, J. W., High-Speed Electrostatic Storage, Annals of the Com- 
putation Laboratory of Harvard University, Vol. XVI, pp. 125-129 (1948). 

63. Haeff, A. V., The Memory Tube and Its Application to Electronic 
Computation, Mathematical Tables and Other Aids to Computation, Vol. 
Ill, No. 24, pp. 281-286 (October, 1948). 

64. Haeff, A. V., A Memory Tube, Electronics, Vol. 20, No. 9, pp. 80-83 
(September, 1947). 

65. Williams, F. C, and T. Kilburn, A Storage System for Use with Binary- 
Digital Computing Machines, Proceedings of the Institution of Electrical 
Engineers, Part III, pp. 81-100 (March, 1949). 

66. Unsigned, An Image Storage Tube, Electronics, Vol. 21, No. 5, pp. 
132-134 (May, 1948). 

67. Zworykin, V. K., and G. A. Morton, Television (John Wiley & Sons, 
Inc., New York, 1940), p. 326. 

The Selectron 

68. Rajchman, Jan A., The Selectron — A Tube for Selective Electrostatic 
Storage, Annals of the Computation Laboratory of Harvard University, 
Vol. XVI, pp. 133-135 (1948). 

69. Rajchman, Jan A., The Selectron — A Tube for Selective Electrostatic 
Storage, Mathematical Tables and Other Aids to Computation, Vol. 2, 
pp. 359-361 (October, 1947). 

70. Rajchman, Jan A., The Selectron — A Tube for Selective Electrostatic 
Storage, RCA Laboratories Division News, Vol. 6, No. 2 (August, 1949). 
A paper presented at the IRE Tube Conference at Princeton University 
in June, 1949, and also at the Symposium on Large-Scale Digital Com- 
puting Machinery at Harvard University in September, 1949. 

Transfer in Space 

71. Bode, Hendrick W., Network Analysis and Feedback Amplifier Design 
(D. Van Nostrand Company, Inc., New York, 1945). 

72. Hansen, W. W., Maximum Gain Bandwidth Product in Amplifiers, 
Journal of Applied Physics, Vol. 16, pp. 528-534 (September, 1945). 

73. Page, C. II., Digital Computer Switching Circuits, Electronics, Vol. 21, 
No. 9, pp. 110-118 (September, 1948). 



384 PHYSICAL COMPONENTS AND METHODS 

74. Rubinoff, Morris, An Input Device Using Multiple Gates, Annals of the 
Computation Laboratory of Harvard University, Vol. XVI, pp. 254-259 
(1948). 

75. Terman, F. E., Radio Engineering (McGraw-Hill Book Company, 
Inc., New York, 1947). 

76. Wheeler, H. A., Wide Band Amplifiers For Television, Proceedings of the 
Institute of Radio Engineers, Vol. 27, pp. 429-438 (July, 1939). 

Coincidence Circuits 

77. Bay, Z., and G. Papp, Coincidence Device of 10 -8 — 10~ 9 Second 
Resolving Power, Review of Scientific Instruments, Vol. 19, No. 9, pp. 
565-567 (September, 1948). 

78. Dickie, R. H., A High Speed Coincidence Circuit, Review of Scientific 
Instruments, Vol. 18, pp. 907-914 (December, 1947). 

79. Hofstadter, R., and J. Mclntyre, Note on the Detection of Coincidences 
and Short Time Intervals, Review of Scientific Instruments, Vol. 21, No. 
1, pp. 52-53 (January, 1950). 



CHAPTER 15 
DATA-CONVERSION EQUIPMENT 

15-1. Introduction 

The essence of the computing problem may be said to be the 
expedient handling of data from origin to final state. However, 
the steps between the original sensing of values and their final 
presentation after a calculation are sometimes complex and may 
require considerable time. Moreover, it may be desirable to 
accept original data in one form, say as decimal digits, perform 
computations on these data expressed in another form, say as 
binary digits, and produce results in still a different form, say 
as graphs. 

Let us consider all the steps in the conversion of data from 
their primary state into an appropriate form for output from a 
computer. 

The sources of data which are suitable for automatic com- 
puting and the channels through which the data may flow to the 
computer are highly diverse. For example, machines used by 
the Bureau of the Census handle data which were passed by word 
of mouth to a census taker and jotted down by him in a note- 
book. The data for a Fourier analysis of a complex wave form 
may come from an electromechanical pickup mounted on an 
aircraft engine. Number-theory problems often originate in a 
mathematician's mind. 

Data may accumulate rather slowly from a number of different 
sources, as in the census example, in which case the data must be 
stored and transported to a central location. Data may emanate 
from one source and accumulate rapidly, as in the engine-vibra- 
tion example. Here the computations may be performed at the 
time and place where the data are being sensed, and intermediate 
storage of data may be avoided. The latter arrangement has 
been termed on-line operation. 

It may be desirable to convert the data from analog form to 
digital form, or vice versa. For example, it may be necessary 

385 



38G PHYSICAL COMPONENTS AND METHODS 

to use as one of the problem inputs the instantaneous value, at 
1-second intervals, of an output voltage which varies with time. 
The instantaneous values of this voltage, appearing as physical 
voltages and not as numbers of volts, are analog data. Measuring 
a physical quantity and expressing it as a number (of volts, 
inches, degrees, etc.) constitute conversion from analog to 
digital form. Analog computers operate directly on analog 
data; digital computers operate on numbers, as was explained in 
Sec. 11-1. 

When data are introduced into a high-speed machine, it may 
be desirable to enter them at a faster rate than operators can 
handle. Therefore, speed-changing devices may be required to 
introduce data rapidly. Likewise, computed results may be 
produced more rapidly than the output medium can assimilate 
them. In this case, the results must accumulate in some inter- 
mediate form until they can be written in final form. Finally, 
when data are written, they should be in a form which will be 
convenient for the user. 

15-2. Analog -to -digital Converters 

An analog-to-digital converter is a device which accepts 
instantaneous values of continuously variable quantities and 
expresses them in discrete numerical form. 

15-2-1. Elementary Types. An example of an analog-to- 
digital converter is the crystal-controlled pulse generator and 
associated ring counter of Sec. 3-6, which measures and expresses 
time intervals in discrete numbers of electrical pulses. 

Another example is a device for converting speed of rotation to 
digital form A cam-operated electrical contactor, fastened to a 
revolving shaft, opens and closes its contact N times for each 
shaft rotation, N being the number of raised segments on the 
cam. Incidentally, this converter has the disadvantage that 
positive and negative directions of rotation are represented in 
the same manner, and there is no zero reference. In other 
words, it counts modulo N but not radix N. 

15-2-2. Meters and Counters. In the familiar watt-hour 
meter 2 there is a set of gears with a ten-to-one reduction ratio, a 
pointer, and a dial numbered from to 9 for each decimal digit 
to be registered. This device is a measuring instrument, as 
distinguished from a counter. The dials rotate smoothly, at a 



DATA-CONVERSION EQUIPMENT 



387 



rate proportional to the instantaneous power usage; the device 
measures the time integral of power usage. At regular intervals 
the meter reader looks at the dials and assigns a discrete value to 
each dial setting. In assigning discrete values to these settings, 
he is performing a conversion of data from analog to digital form. 
In fact, the meter reader himself satisfies the definition stated in 
Sec. 15-2. 

Meters (which yield analog data) may be distinguished from 
counters (which yield digital data) on this basis: a counter is 
arbitrarily precise in its representation of a quantity in digital 

2 TEETH 
INPUT 




STRIPPED W 
GEAR 
DRIVE 



Fig. 15-1, 



DIAL 



DETENT MECHANISM 
Stripped-gear counter. 



form, and no particular skill is required to read it; on the other 
hand, some skill is required to read a meter, because the meter 
measures and expresses a quantity, or, at least, the last digit of 
its measured value, in analog form. The higher the degree of 
precision required in the expression of a particular meter reading, 
the greater the observer's skill must be and the less likely two 
different observers are to read identical digital values from the 
same dial or scale and pointer configurations. 

An example of a typical rotation counter is shown in Fig. 15-1. 
This counter has numbered dials which move in discrete steps 
from one digit to the next. This counter employs a driving 
pinion connected to an input shaft stripped of all but n/r of its 
teeth, where n is the original number of teeth and r is the radix 
of the number system being used. The stripped pinion drives a 
gear having n teeth, possibly through an idler gear, and advances 
it 1/r revolutions for each revolution of the stripped driver. 
Thus on a radix 10 counter, in which the position of the full- 
toothed gear is represented by the digits to 9 around a dial, a 



388 PHYSICAL COMPONENTS AND METHODS 

new decimal digit appears at an aperture after each revolution of 
the input shaft. A single decimal counter is illustrated in Fig. 
15-1. A group of such dials, cascaded so that the output of one 
rotates the stripped gear of the next, form a decimal counter on 
which as many digits as desired may be represented. A detent 
mechanism may be required to hold the full-toothed gears in 
their discrete positions when they are not meshed by the teeth on 
the partially stripped gears. The dials on these counters are 
commonly placed on the outside of short cylinders, which are 
located behind a window aperture and arranged so that only one 
digit on each cylinder is in view at any time. 

Mechanical counters are manufactured by the Veeder-Root 
Company. Various models of this device count rotations, 
reciprocating displacements, and electrical pulses. The output 
of these counters is normally a set of numbers to be read visually. 
However, several models are available with printing wheels as 
an optional feature. The top speed recommended for any 
Veeder-Root Counter is about 42 counts per second. Most of 
them operate at about 15 counts per second maximum. 

Mechanical counters have also been equipped with electrical 
pick-off s at each discrete position in order to generate electrical 
signals representing shaft position. An analog-to-digital con- 
verter which consists of a group of rotary switches geared to a 
rotating shaft is used in the MIT differential analyzer. 1 These 
switches in conjunction with conversion circuits operate an 
electric typewriter and type numbers which indicate the shaft 
location. Another electrically read device is used in the Harvard 
Mark I and is known as an electromechanical counter wheel. 
Each counter wheel has 10 discrete positions and a set of electrical 
contacts for each position. The wheels are driven to their 
prescribed position by means of a common drive motor through a 
magnetic clutch. The clutch actuation is controlled by a 
number-input mechanism which causes each wheel to be driven 
to the digit to be represented. 

15-2-3. Film Converter. Another rotational analog-to-digital 
converter has been proposed which can be arranged to convert 
shaft rotations to numbers of radix 2. This converter employs 
a continuous strip of film on which transparent and opaque spots 
are photographed. These spots are arranged in columns across 
the width of the film to represent binary numbers, which increase 



DATA-CONVERSION EQUIPMENT 



389 



by one unit for each successive column along the length of the 
film. Thus, a 15-digit binary number is represented by a column 
of 15 code areas across the film and provides an opportunity to 
employ up to 2 15 different numbers for the shaft rotation involved. 
The film drive is of the intermittent type used in motion-picture 
projectors and thus fulfills the definition of a continuous-to- 
discrete-value converter. Common types of intermittent drives 
are the Geneva movement and the Powers pin-cross and the 



BANK OF 
PHOTOCELLS 




LOOP OF DIGIT- 
-CODED MICRO- 
FILM 



AMPLIFIERS 


\ 


CONTROL 
GATE 




1 






STORAGE 
UNIT 




OUTPUT 





NPUT SHAFT 



Fig. 15-2. A scheme for converting analog to digital data. The micro- 
film drive moves smoothly while coded areas on the film are sensed photo- 
electrically. 

star-wheel drives. 3 A given amount of input shaft rotation 
causes the film to move forward one discrete frame or, in this 
case, one column of code areas. If a photocell is used in con- 
junction with each of the 15 rows of spots in a column and if a 
light source is arranged to shine through the film and on the 
respective photocells, as described in Sec. 14-2-3, a radix 2 
digital representation of the input shaft position is imposed upon 
the photocells. 

A mechanical film drive requires more than a millisecond to 
step the film from one code column to the next. In order to 
increase the maximum film speed and to eliminate ambiguity 



390 PHYSICAL COMPONENTS AND METHODS 

during the advance time, an electronic synchronizing scheme 
could be used instead of the intermittent film drive. A row of 
synchronizing dots on the film might control gates which allow 
the photocells to be sensed when the code pattern is centered in 
the optical system, but at no other time. Thus, each number 
is held until it is possible to sense the next one correctly. A con- 
verter of the type described has been proposed by ERA and is 
illustrated in Fig. 15-2. It is thought that 50 binary digits can 
be stored across the width of a 35-millimeter film. 

15-2-4. Pulse Code Modulation. A wide-band communica- 
tion system which has desirable signal-to-noise characteristics 
has been devised using a binary representation of analog sig- 
nals. 4-11 This system is called pulse code modulation, or p. cm., 
and is of interest because of the converters which are used and 
because of the possibility of transmitting analog data to a com- 
puter in digital form over a telemeter link. 

Three general types of analog-to-digital converters have been 
devised for p. cm. In all three schemes, the instantaneous 
amplitude of the analog signal is measured or sampled periodi- 
cally. The sampling rate is made slightly more than twice the 
highest frequency component which is desired to be sent. Higher 
frequency components are filtered out of the signal previous to 
the sampling procedure. 

In the first type of p. cm. converter, the number of voltage 
units in an amplitude sample are counted successively a unit 
at a time until a residue of less than one unit remains. 4,7 In the 
second type of p. cm. converter, the amplitude of the sample is 
compared with a standard voltage which corresponds in mag- 
nitude to the highest order digit, and the standard voltage is 
subtracted if it is smaller than the sample. The remainder of 
the sample is then compared with a voltage corresponding to the 
next highest order digit, and another subtraction is made if the 
sample is larger. This process is continued until the lowest 
order, or unit, digit is compared with the remainder and is deter- 
mined to be either larger or smaller. 6 In the third type of p. cm. 
converter, the sample amplitude is measured, and the correspond- 
ing digit pulses are produced without intermediate counting or 
subtracting. 8,1011 The speeds of these three systems increase 
in the order of listing given above. 

Examples of the first and third converters are given in order 



DATA-CONVERSION EQUIPMENT 



391 



to illustrate their operation and circuitry. The first is perhaps 
the most simple of the p. cm. conversion schemes and will serve 
as a basis for comparison in describing the third type, which is 
the fastest. 



INPUT GENERATOR 



SIGNALS 




-0- 

I PULSE DELAY 

Fig. 15-3. Pulse code converter. The instantaneous amplitude of a signal 
produced by the input generator is periodically sampled and converted into 
a proportionate number of clock pulses by the modulator and gate and into 
digital representation by the counter. This number of pulses must be 
between and 15 in the equipment illustrated. Here two samples of the 
signal are being taken; the amplitudes at the points where the signal is 
being sampled, expressed in numbers of pulses, are shown as 10 and 5 (with 
respect to some bias voltage). 

The first scheme was patented by A. H. Reeves and has been 
investigated by Bell Telephone Laboratories and by the Federal 
Radio and Telephone Corp. 4,7 A sample of the voltage represent- 
ing the data to be converted is first introduced into a sampling 



392 PHYSICAL COMPONENTS AND METHODS 

pulse-width modulator which measures the amplitude of the 
incoming voltage with respect to a reference bias voltage and 
gates out a pulse having a width proportional to the measured 
amplitude. The variable-width pulses control a gate in such a 
way that a number of clock pulses nearly proportional to the 
width of the pulse is passed through. If the analog data are 
to be represented by four binary pulses, they are represented to 
within 1 part in 2 4 , or 16. Therefore the clock-pulse rate is 
set so that pulses of maximum width consume the same amount 
of time as 16 clock intervals, and each amplitude sample is 
replaced with from to 15 clock pulses which represent the 
amplitude of the sample. The series of clock pulses is counted 
in a four-stage binary counter of the type described in Sec. 13-5 
on ring counters. A block diagram of this system is illustrated 
in Fig. 15-3. 

The third type of p. cm. conversion has been accomplished 
at the Bell Telephone Laboratories 8,10 ' 11 using a perforated-mask 
type of converter. The mask intercepts an electron stream in a 
tube similar to a conventional cathode-ray tube. The beam 
passes between two perpendicular pairs of electrostatic deflecting 
plates which serve to locate the beam on the mask. A positive- 
potential pulse plate is mounted at the large end of the tube 
perpendicular to the longitudinal axis, and the mask is. placed 
in front of it. Thus, the electron stream may or may not reach 
the collector plate, depending upon whether it is deflected to 
pass through a hole in the conducting mask or not. The collector 
plate might be replaced with a photoelectric cell which detects 
the presence of a light spot anywhere on a conventional c.r.o. 
screen. 

The analog voltage is sampled periodically, and a voltage 
which is proportional in amplitude to the sample is applied to 
one set of deflection plates. Simultaneously, a sweep voltage is 
applied to the other pair of deflection plates so that the electron 
beam scans across some portion of the mask. The mask has 
holes cut in it similar to the pattern illustrated in Fig. 15-4 so 
that if the sample amplitude controls the vertical location of the 
beam, the electrons reaching the collector will be modulated by 
the mask in a manner dependent upon the code pattern. The 
code pattern illustrated represents the binary numbers from 
to 15, and consequently the electrical pulses on the collector 



DATA-CONVERSION EQUIPMENT 



393 



plate produce a serial binary number which is roughly propor- 
tional to the analog sample. Auxiliary equipment is used to 
assure that the electron beam 
moves along the center of a row 
of code areas and not along the 
edge. 

This device has been used to 
sample 12 channels of voice fre- 
quencies up to 3,400 cycles per 
second simultaneously utilizing 
a mask which is seven binary 
digits wide. Electrical pulses 
1.5 microseconds wide are trans- 
mitted at a rate of 672 kilocycles 
per second. 



15-3. Digital-to-analog 
verters 



Con- 




Fig. 15-4. An analog-to-binary- 
converter screen. 



After the required computa- 
tions have been performed in a 
digital computer, it is sometimes 
desired to transform the digital 
results into analog form for ease 
of viewing and assimilation. 
Three types of digital-to-analog conversion are discussed here: 
conversion of pulses to voltage amplitudes, conversion to rotary 
motion, and conversion directly to a graphical representation. 

15-3-1. Pulse Converters. A series of electrical pulses may be 
transformed into a d-c voltage by allowing them to gate stand- 
ardized pulses into a capacitor having very low leakage. The 
voltage across the terminals of the capacitor at any instant is 
proportional to the number of pulses gated into the capacitor 
since the last time it was fully discharged. This transformation 
method was the basis of some of the early measuring circuits 
associated with Geiger-Muller counters and is also used in 
frequency dividers. 

A digital-to-analog converter employed in the Bell Telephone 
Laboratories p. cm. system makes use of the exponential decay 
characteristics of an RC circuit. The digits of a binary number 
are represented by standardized electrical pulses, separated 



394 



PHYSICAL COMPONENTS AND METHODS 



evenly in time and arranged so that the lowest order digits 
appear first. They are applied to a capacitor circuit with the 
RC constant so adjusted that half the charge leaks off between 
pulse times. The total charge after all the digits in a number 
have been inserted is proportional to that number. 

For example, consider a converter in which four-digit binary 
numbers are introduced in pulse times 1 to 4 and are removed 
at pulse time 5. If the binary numbers 1,010 and 1,111 are 
converted in this fashion and inserted pulses are 80 volts in 
amplitude, the voltage on the capacitor as a function of pulse 
times is shown in Table 15-1. 



Table 15-1 



Pulse 


Binary number = 1,010 
Decimal number = 10 


Binary number = 1,111 
Decimal number = 15 


time 


Binary 
digit 


Capacitor 
volts 


Binary 
digit 


Capacitor 
volts 


1 
2 
3 
4 
5 




1 

1 


0+0 =0 

+ 80 = 80 

40 + = 40 

20 + 80 = 100 

50 


1 
1 
1 
1 


+ 80 = 80 

40 + 80 = 120 

60 + 80 = 140 

70 + 80 = 150 

75 



It can be seen from the above example that the decimal 
conversion factor is 5 volts per unit at the pulse time. The 
converter described was proposed by Shannon at Bell Telephone 
Laboratories. An important modification has been developed 
by Rack 5 which makes timing less critical. 

15-3-2. Rotation Converters. By a slightly different arrange- 
ment the rotational analog-to-digital converters previously dis- 
cussed can be operated in an inverse manner to produce a rota- 
tional output for a digital input. A servomechanism 12-15 may 
be coupled to the film converter or electromechanical wheel 
counter in such a way that the motor drives the shaft until a 
desired number is located. 

As an example, consider the film converter of Fig. 15-2. 
Numbers from to 2 4 — 1 are represented successively in binary 
form as transparent and opaque dots . on a four-row field as in 



DATA-CONVERSION EQUIPMENT 395 

Fig. 15-4. The fourth-order or left-hand row of digits is half 
O's and half l's (eight of each). The photocell output for that 
row is compared with a signal representing the desired digit, 
and the servo drives the film in the correct direction so that the 
fourth-order digits are caused to match. After the servo has 
driven the film to the nearest point where fourth-order digits 
match, a comparison of third-order digits is made. Within the 
limits in which fourth-order digits are matched, half the third- 
order digits are l's and half are O's (four of each). Again, the 
desired third-order digit is compared with the third-order digit 
on which the film is set, and the film is either left where it is or 
driven to the first point where third-order digits coincide. This 
process is continued until the film is driven to a point where its 
digits coincide with those of the number desired. 

Another example utilizing the principles involved in the film 
converter can be found in the MIT differential analyzer. 1 This 
is a tape-operated device which sets the displacement of the 
integrator disk to within 1 part in 150,000 of the maximum 
displacement after a punched tape is prepared by an operator. 
The conversion from the digital value on the tape to shaft rota- 
tion is accomplished with the aid of 44 relays and a set of rotary 
selector switches on the output shaft. The shaft is driven by a 
servomotor which, in turn, is controlled by the switches. First 
the relays are set by electrical contactors actuated by the holes 
in the punched tape; then the servomotor is caused to rotate in 
the direction to close corresponding switches on the integrator 
shaft. The motor continues to drive the integrator shaft until 
a switch corresponding to each digit is closed, whereupon it is 
caused to stop. 

15-3-3. Use of the Selectron as a Digital-to-analog Converter. 
As originally conceived (see pages 370-371) the Selectron had a 
64-by-64 grid, on a cylindrical surface. By deforming such a 
grid into a plane, it would be possible to produce a device for 
converting digital data directly into graphic form. The electron 
beam would be directed upon a spot on the screen opposite a par- 
ticular window in the grid for each digital value; the composite 
of the spots would form a graph. 

For example, if some digital F{x) is to be plotted as a function 
of digital quantities, represented by the letter x, the x quantities 
are caused to operate the gates associated with the vertical bars, 



396 PHYSICAL COMPONENTS AND METHODS 

and the F(x) quantities are simultaneously applied to the hori- 
zontal bars. Thus, in a Selectron system having 2 5 vertical bars, 
values of x ranging from to 2 5 — 1 might be plotted. Then, 
all values of x between and 1 would cause the left-hand set of 
bars to be made positive; all values between 1 and 2 would cause 
the next pair to the right to become positive, etc., so that values 
between 30 and 31 would cause the right-hand pair of bars to 
become positive. A similar scheme would be used to introduce 
F(x) values on the horizontal bars, and quantities would be 
plotted to within an accuracy of about 3 per cent. If x and 
F(x) values are directed to the vertical and horizontal bars 
simultaneously, a Selectron window is opened at a point on the 
grid which represents the value of these two quantities. Num- 
bers to be plotted can be brought into the range of the plotting 
device by a previous multiplication or division by a scale factor 
as required. A photosensitive screen placed behind the Selec- 
tron grid, in place of the dielectric storage plate, would cause a 
visual spot to be obtained for each x versus F(x) entered. 

15-4. Conversion of Data from Written to Automatic Machine 
Form 

Data for automatic computation often are presented in a 
written form not easily assimilated by a computer. They must 
be converted therefore into some new form for introduction 
into an automatic machine. 

The most common written form for recording data is the famil- 
iar numerical tabular form used in accountants' records, engineer- 
ing data sheets, and tables of mathematical functions. Data 
may be recorded also as a set of plotted points, which represent 
in analog form a function of one parameter in terms of discrete 
values of another parameter. As a variation of this form, a set 
of lines may be drawn from the plotted points perpendicular to 
one coordinate axis. These two presentations of plotted data 
will be referred to as sampled analog data. 

The numerical tabular form of written data may be introduced 
into a computing machine by a human operator who reads the 
data and manipulates dials or a keyboard in a manner which 
corresponds with his mental evaluations of the data. Sampled 
analog data may be transferred by an operator in one of two 
different ways. First, the various points of a plotted set may 



DATA-CONVERSION EQUIPMENT 397 

be measured from a reference with a scale or ruler. Numerical 
measure thus is afforded which may be introduced into a machine 
in a manner similar to numerical tabular data. A second method 
of transferring plotted data is to superimpose a sighting device 
upon the points, thereby producing shaft displacements which 
may be used in the machine as measures of the function values. 
This method permits the transfer of data without requiring the 
operator actually to sense the values mentally. 

Data in the form of continuous curves may be evaluated in 
a manner similar to the methods used for sampled data. Auto- 
matic curve-tracing devices have been made which utilize photo- 
electric-cell sensing to control a motor drive for the curve follower. 

In this section are included descriptions of equipment which 
has been devised to perform data conversion from a written form 
into a form suitable for introduction into an automatic computing 
machine. In addition, a description is included of a proposed 
device with which data may be entered into a computer from 
vocal commands. 

15-4-1. Conversion from Written to Automatic Digital Form. 
The automatic digital forms into which data may be transcribed 
are discussed in Sec. 14-2 under the topic of storage mediums. 
The ways in which data have been transcribed from written form 
to these automatic digital forms are numerous, but all employ 
the idea of the keyboard or the dial. It is important that the 
converting process allow the operator to keep his eyes on the 
data being read. 

Keyboards. There are several types of keyboards in common 
use. The simplest of these is the single key used by the tele- 
graph operator. The semiautomatic key for high-speed teleg- 
raphy has a neutral position, a position for dots, and a position 
for dashes so that it can be used to generate binary numbers 
serially. 

The Sundstrand keyboard is used on some desk calculators. It 
employs only 10 keys representing the 10 decimal digits. These 
keys are depressed successively, and a number is formed in a small 
register window where it can be checked before the operation 
begins or before the number is transferred to the storage medium. 
This keyboard can be operated quite rapidly by the touch 
system. 

The Burroughs keyboard, in which lOn keys are required to 



398 PHYSICAL COMPONENTS AND METHODS 

set up an n-digit number, is common on desk computers. It may 
be operated by the touch system, but not as rapidly as the 
Sundstrand. The desirable feature of this keyboard is that after 
all the numbers are set on the keyboard, they may be checked 
visually, and any errors may be corrected easily before the 
computation with the number begins. 

A familiar keyboard is that of the ordinary typewriter, which 
may be operated at speeds of about five letters per second. It 
has the desirable feature that many operators are already pro- 
ficient in its use. A keyboard based on the standard typewriter 
keyboard is being used as the input device for the UNIVAC. If 
a typewritten record is made in conjunction with the data-con- 
version process, it may be checked for errors, which may then be 
corrected before computations begin. 

Dvorak 18 has suggested rearranging the positions of the keys 
on the typewriter keyboard so that easier finger motions are 
required in typing the frequently used letters and combinations 
of letters. Machines have been equipped with Dvorak key- 
boards and used successfully in speed tests and other demon- 
strations. It is said to be difficult to retrain an experienced 
operator to use this keyboard, but novices can probably learn 
to use it at least as easily as a conventional keyboard. Type- 
writers equipped with this keyboard are available. 

The keyboard on the teletypewriter basically is similar to that 
of the typewriter but has three rows of keys rather than four. 
All letters are upper-case, and numbers are accessible by use of 
the shift key. The machine has a top speed limit of about six 
letters per second. Teletypewriter equipment is used in the 
input equipment for the Mark II and Bell Telephone Laboratories 
machines. 

Another keyboard which has been devised for high-speed 
typing is manufactured by the Stenotype Company. This 
keyboard contains 22 keys which represent the common sounds 
of the language used. The operator uses the fingers of both 
hands and by depressing several keys simultaneously produces 
the associated sounds in the desired order. Stenotype operators 
can attain a speed of 250 words per minute. Part of this increase 
in speed results from the reduced number of keys which must be 
struck for each word written, and part of it results from the fact 
that several keys may be depressed simultaneously. Although 



DATA-CONVERSION EQUIPMENT 399 

this machine is not particularly suitable for use in a computer, 
the possibility of using the principle of making up numbers and 
orders out of combinations of keys might well increase the speed 
of transcription from written to automatic machine forms. 

Keyboard Coupling Devices. Manually operated keyboards 
are usually employed to record data on storage mediums for 
future use in a computer. The actual coupling between the keys 
and the recorder unit then depends to a large extent upon the 
type of keyboard and the type of storage employed. There are 
possible a large number of coupling devices, some of which are 
suggested in this section. The teletypewriter is described as an 
example which employs mechanical coupling, electric switches, 
and an encoding commutator. 

When storage mediums such as punched tape and punched 
cards are used, the coupling between keyboard and punches may 
be entirely through mechanical linkages. These linkages are 
arranged to set the proper punches, which are usually driven 
through the paper by an auxiliary power source. 

When an electrical output is desired from the keyboard, the 
keys may operate switches directly or may operate small shutters 
which control a supply of light to a group of tiny photocells. The 
latter principle is applied in Teleregister Corporation equip- 
ment. 16 Photocells have the advantage over switches that they 
do not have contacts which may become dirty and thus cause 
intermittent operation. Other couplings, which might be 
devised to overcome this rather fundamental difficulty with 
small switches, are electrostatic and inductive devices which 
can be coupled and uncoupled with a keyboard. 

The teletype machine is a combination keyboard and printer. 20 
The keyboard section contains five switches, and each key causes 
one or more of these switches to be closed when it is depressed. 
The coupling from key to switch is through levers, cams, and 
five sets of push rods, which run the width of the keyboard. 
When a teletypewriter is used in conjunction with a tape per- 
forator, the five switches control five pin-setting devices. An 
electromagnet in the punching machine causes the set pins to be 
driven through the paper tape. Thus, a unique hole pattern is 
produced in the tape for each key depressed. 

When the output of the teletype machine is to be transmitted 
by a single wire, the five switches are connected to the central 



400 PHYSICAL COMPONENTS AND METHODS 

five positions on a seven-place electromechanical commutator. 
When a type key is depressed, corresponding switches are closed 
and latched, and the commutator arm begins a single rotation. 
In its first position a start pulse is placed on the single wire; in 
the next five positions positive or negative pulses corresponding 
to the particular key depressed are generated; and in the seventh 
position a stop pulse is sent out. At the completion of the 
cycle the switches are released, and the machine is ready for the 
next key to be depressed. Thus, for each letter typed on the 
keyboard, a string of seven positive and negative pulses are 
emitted serially from the single-wire transmitter. 

The receiver unit of a teletypewriter consists of a mechanical 
amplifier which increases the mechanical power available from 
the relatively weak incoming signals so that they can be separated 
and used to swing a type bar corresponding to the key depressed. 

When a teletypewriter is set to send or receive at one time, the 
receiver unit may be electrically connected to the output of the 
keyboard so that a copy of the material being sent out is typed at 
the sending station. This arrangement aids the operator in 
keeping track of lines and in locating errors. Since the keyboard 
and the printer are located together, one unit suffices for sending 
and receiving. 

The IBM Corporation produces the Electromatic Typewriter, 
which is available on a rental basis. Although this machine was 
not originally intended for transmitting information, the Globe 
Wireless Company has modified these machines so that they are 
suitable for this purpose. In the IBM machine the keys operate 
combinations of six switches simultaneously. The Globe 
Wireless Company has added an electronic conversion unit which 
transforms the information in six static switch settings into a 
serial string of six pulses. When the key is struck, it also causes 
a type bar to strike the paper in the manner of a conventional 
mechanical typewriter, so that a record of the message sent is 
also made at the transmitter. 

A separate machine is required as a receiver in the Globe Wire- 
less equipment. This machine has an electronic converter called 
a permutation unit, which distributes the six serial pulses into six 
channels. Each of these six channels controls an electromagnet, 
which in turn operates code levers. These code levers operate 






DATA-CONVERSION EQUIPMENT 401 

the type bars which actually print the characters corresponding 
to the serial pulses received. 

Another tape-punching typewriter designed for special uses 
in offices is made by the Justowriter Company. It, also, is 
available on a rental basis. 

Dials. A second conceivable manner in which an operator 
might transcribe data from numerical form to automatic digital 
form is by setting the numbers on dials with discrete positions 
on them. These positions might be determined by a detent 
mechanism arranged to hold the dial on positions corresponding 
to whole numbers. A variation of this idea is illustrated by the 
telephone dial, with which discrete numbers are registered 
by means of discrete holes in the finger ring. 

Speech Converters: One of the most common forms of error 
which operators make in keying or dialing numbers is to transpose 
digits. Furthermore, a certain amount of skill is required in 
order to operate a keyboard speedily and accurately. An 
input device which allows the operator to concentrate on the data 
which he is entering into the machine and which requires little 
practice to use might be devised from a speech converter. The 
operator merely reads the digits of a number into a microphone, 
and the digit names are recognized by selective filters in an 
amplifier. On the basis of this recognition, gates are opened 
which convert the sound of the spoken digit into automatic 
digital form. The reproduction of speech from elementary 
sounds was studied at Bell Laboratories prior to 1938 17 and has 
recently been reconsidered in England. 19 

15-4-2. Conversion from Written to Automatic Analog Form. 
A simple way in which to convert numerical values into analog 
form suitable for automatic machine operation is by means of a 
dial mechanism or keyboard. Plotted data may be converted 
for introduction to a computer either by measuring and entering 
from numerical values or by sensing with a curve-following 
device connected to a displacement mechanism. 

Dials, Scales, and Keyboards. A dial may be used to establish 
an angle of rotation on a shaft, to set a potentiometer, or to 
rotate a synchro. 20 A scale is customarily used for measuring 
linear distances. The familiar slide rule, for example, has 
linear scales along its edges, and numbers may be converted from 



402 PHYSICAL COMPONENTS AND METHODS 

digital to analog form and back again by the operator several 
times in the course of an arithmetic operation. 

Similarly, analog values could be set with a keyboard in a 
number of different ways. For instance, each key might operate 
a lever arm having a length proportional to the key number. 
Also, each key might operate an electrical contactor and thereby 
switch impedances or voltages in a manner designed to create 
appropriate analog values. Another possibility is to have keys 
set stops in some device which moves until it hits a stop, thus 
converting key settings to linear or angular displacements. 

A keyboard is used originally to punch tape with which the 
integrator disks are set on the MIT differential analyzer as 
described in Sec. 15-3. 

Curve Followers. An operator can transfer discrete or con- 
tinuous values plotted on paper into automatic form by means 
of a curve follower. In essence, the operator controls a cross 
hair or similar sighting device so that it moves from one to 
another of a set of discrete plotted points or along the curve of a 
continuous record. The position of the tracing element with 
respect to a selected reference point corresponding to the origin 
of the curve becomes the analog output of the machine, which 
may be in the form of shaft rotations or linear displacements. If 
a curve is plotted in rectangular coordinates designated by x 
and y, the medium on which the curve is drawn may be moved 
by a motor drive in the x direction. The operator then controls 
the y position of his sight and possibly the speed and direction 
of the x drive. Such a scheme is easier for the operator to use 
than one in which he controls both axes completely and simul- 
taneously. Equipment requiring manually operated drives 
similar to the one just described is common in gun-laying and 
bombing computers. 

The output of a curve follower usually consists of a pair of 
shaft rotations which are proportional to two plotted quantities. 
Sometimes one shaft will be driven by another part of the 
machine, in which case only the operator-controlled shaft will 
constitute an output. As a check upon the accuracy of the 
operator, a pencil can be attached to a curve follower so that the 
distance between the recorded curve and the curve traced out 
by the operator can be determined. 

Automatic curve followers have been considered for some time, 



DATA-CONVERSION EQUIPMENT 403 

and one was built by Hazen, Jaeger, and Brown at MIT and 
reported in 1936. 22 It is capable of following a single-valued 
function of one variable within the range where its first derivative 
remains finite. Their machine was successful and is the fore- 
runner of a similar automatic curve follower which was developed 
in Germany at the Technische Hochschule, Darmstadt. 23 It 
operates on the basis of a photocell which rides along one edge 
of a black line on a light background. An image of an illuminated 
section of the line is focused on the photocell through a circular 
aperture which passes a spot of light having a diameter equal to 
about one-third the width of the line. The equilibrium position 
of the photocell is established at the place where half the image 
is black and half the image is white. With this arrangement an 
increase in illumination, which results when the photocell is 
moved away from the center of the line, energizes a servomecha- 
nism, which drives the photocell back toward the center of the 
line, and a decrease in illumination causes the photocell to be 
driven away from the center of the line. Since the servodrive 
is along the y axis, the sight is actually driven in the -\-y direction 
in one case, and in the — y direction in the other. Therefore, 
it is essential that the sight remain on the side of the curve on 
which it will always be led back toward the center of the line. 

The accuracy of this curve follower is a function of the amount 
of hunting which occurs in the servosystem. The amount of 
hunting depends in a rather complicated manner upon the circuit 
parameters and the input functions and is described in the 
theory of servomechanisms. 12,14 ' 15 In general, the inertia of 
the system should be as low as possible to reduce hunting ampli- 
tude. Increased damping decreases the hunting but also 
increases servo lag and consequently error. 

With a given servosystem and a limit set on the error which 
can be tolerated, the velocity of the servo movement must be 
restricted. This velocity is equal to the product of the x-direc- 
tion velocity of the curve times the slope of the curve. In the 
German machine a speed of 3 millimeters per second in the x 
direction is a safe drive speed. The error caused by the aperture 
hunting a distance d normal to the curve and with the curve 
making an angle a from the y axis is 

d 



E 



COS a 



404 PHYSICAL COMPONENTS AND METHODS 

and on the German machine a value of d — 0.1 millimeter was 
attained. 

Other automatic curve followers have been considered in which 
a second photocell is used to detect the slope of the curve. The 
main servo-circuit parameters thereby may be controlled in such 
a way that the curve may be followed efficiently over the wide 
range of required servo velocities resulting from the wide range 
of slopes on various parts of the curves to be followed. An 
application of modern servomechanism circuit techniques to this 
problem would probably be well rewarded. 

The automatic curve follower just described, like the MIT 
device, is limited to sections of curves in which the first derivative 
or slope is finite, since the control circuit is unstable when its 
orientation from the center of the curve is reversed. Manually 
operated followers, however, are not limited in this respect. 
When a curve follower is guided by a human operator, any of 
several curves plotted on the same medium can be chosen and 
even changed during the course of the problem if data are 
changing slowly. Therefore, functions of more than one variable 
can be followed manually. Use of an additional parameter is 
not feasible with an automatic curve follower, since it is unable 
to distinguish one curve from another or to interpolate between 
curves. 

15-5. Transcription of Data from Automatic Machine to Written 
Form 

After a computation is completed, it may be desirable to study 
the results, and therefore the output of the machine must be 
produced in a written form. This is not particularly necessary 
in a gun-director computer or similar tactical device, where the 
results are used as they are computed, but it is important in 
solving the problems of science. The written values may take 
any of several forms, the most convenient depending upon the 
nature of the data and their eventual use. If large amounts of 
data are to be listed in tabular form, it is well to list tabular sum- 
maries of the data also. In some cases it will be most convenient 
to have the output data plotted in graphical form. In this 
section is discussed the conversion of quantities from machines to 
written numerical or to written analog forms. Equipment which 
performs this function is referred to as output equipment. 



DATA-CONVERSION EQUIPMENT 405 

15-5-1. Transcription from Automatic Digital to Written 
Numerical Form. The mechanical, electronic, and photographic 
printers which have been devised for numerical transcription 
are discussed in this section. 

Electric Typewriters. There are several commercially available 
electric typewriters which are of interest as output devices for 
printing results in numerical form. 

The teletype equipment, 20 described in Sec. 15-4-1, is a com- 
bination keyboard and printer. It can be used to type numbers 
and letters from groups of synchronized pulses. Similarly, the 
modified IBM Electromatic Typewriter and the Justowriter 
equipment can be operated from punched tape or other sources 
of electrical input signals. RCA Communications, Inc., also 
produce a number of tape-operated typewriters. 28 

In a conventional typewriter, when a key is depressed, an arm 
with an individual type face on it is caused to fly up and hit the 
paper, thereby printing a letter or a number. There is another 
group of typewriters in which the font of characters is placed on a 
small wheel located behind the paper. When a key is struck, 
the wheel is caused to rotate until the corresponding type lies 
behind the paper, which is then struck by a hammer from the 
side opposite the font. This is the principle of the early Ham- 
mond typewriter, a Dow Jones machine, a Western Union 
machine, and the Vari-typer. The action is simple, and it is 
rather easy to change type fonts. Typewriters are essentially 
serial printers since only one letter is printed at a time. 

Multibar Printers. A somewhat faster device is known as the 
multibar, or gang, printer in which all the type is arranged along 
a bar, and several like bars are mounted side by side. In one 
version of the multibar printer, each bar is displaced longi- 
tudinally so that the correct figures are aligned to form an anvil 
behind the paper. Letters and numbers are printed by means of 
a hammer which hits the paper and presses it against the bars. 
As an example, a printer having three bars each containing the 
digits to 9 can print three-digit decimal numbers. If the 
number 281 is to be printed, the left-hand bar rises three places 
to the digit 2, the second rises to the digit 8, and the third rises 
so that the digit 1 is under the hammer. Since these bars may 
be raised simultaneously, the total time required for printing is 
the time required to set a bar and to strike it with the hammer. 



406 PHYSICAL COMPONENTS AND METHODS 

A mechanism of this type is found in the ordinary desk adding 
machine. 

In a Remington Rand multibar printer, each letter is mounted 
at the end of a rod which is free to slide longitudinally. A stack 
of these rods is arranged in a carriage which slides up and down 
like the bar in the multibar printer. In this type, the hammer 
hits the correct type rod and presses it against the paper. 

Gang printers which are now available include the Remington 
Rand and IBM printers, which attain a printing speed of over 
100 characters per second. These machines print letters as well 
as numbers. An IBM printer which handles numbers only will 
print at a rate of about 200 digits per second. It has been 
estimated that gang printers with bars actuated by pneumatic 
or hydraulic instead of electromechanical means could be made 
to operate at a rate of several thousand characters per second. 

A gang printer in which the type is located on rotating drums 
rather than bars has been suggested. By rotating the drums 
at a constant velocity, the need to accelerate the type is elimi- 
nated. Printing is accomplished by causing a hammer to strike 
the paper against the wheel as the proper letter is passing by. 
Since the hammer has to move only a short distance, it can be 
brought in place very quickly so that the operating speed of such 
a gang printer should prove to be quite high. 

Flash Photographic Printers. If extremely high speeds are 
required in an output printer, photographic techniques can be 
exploited. They are limited in top speed by the maximum rate 
at which reasonably intense light sources can be modulated. At 
the present time this limit is around 1,000 flashes per second 
with small Edgerton-type flash tubes employed in commercial 
stroboscopes and mentioned on page 320. A compromise 
between flashing rate and illumination is necessary to prevent 
overheating when a flash tube is used repetitiously. Several 
models of this tube are produced by Sylvania Electric Products, 
Inc., and the General Electric Co. The latter is experimenting 
with a flash tube which promises to be smaller and faster than 
the coil-type lamps. 

The Numero scope. A high-speed, photographic recording 
device called the Numeroscope 2b ' 26 for use with a computer has 
been developed at Harvard. Each digit in a number is formed 
on the face of a cathode-ray tube, and a picture is taken of all 



DATA-CONVERSION EQUIPMENT 407 

the tubes in a row. In the system proposed, a 1-inch cathode- 
ray tube will be provided for each of 20 decimal digits. A 
group of 10 generators produce the electrical wave forms required 
to produce each Arabic numeral on a cathode-ray- tube screen, 
and these signals are gated to the proper tubes by the output of 
the computer. The 20 tubes are arranged side by side in a light- 
proof box with a shutterless camera which is aimed and masked 
to photograph a row of numbers across the cathode-ray tubes. 
After a frame of numbers is photographed and has disappeared 
from the screens, the film is advanced a short distance in readiness 
for the next exposure. 

Most of the work reported on the Numeroscope consisted in 
the development of the electronic circuits required to form the 
figures. It was decided after an analysis that it would be very 
tedious to generate the figures with a combination of the required 
sinusoidal components; hence nonlinear circuits were devised 
with which to operate on sine waves of 2,000 and 5,500 cycles 
per second. Each figure is broken down into straight lines and 
sections of ellipses which are generated in turn and correctly 
positioned on the face of the tube. For example, the figure 8 
is a small ellipse which is alternately moved up and down so 
that it appears as one ellipse on top of another. Half-wave- 
rectified signals are used to produce the figure 7, one leg of which 
is produced during one half of the cycle and the other leg during 
the alternate half. 

The smallest figure which can be photographed is a function 
of the graininess of the film, and the exposure time is a function 
of the emulsion speed. The speed of this device in number of 
digits which can be photographed in a second can be determined 
from a knowledge of these two factors. Recent information 27 
indicates that writing speeds of 270 inches per microsecond can 
be photographed on a P-ll screen with high accelerating poten- 
tials using an //1.5 lens and E. K. Linagraph Pan (5244). 

ERA Rapid Printer. At the time of this writing, ERA is 
developing a high-speed printing device under contract with the 
National Bureau of Standards. In this device, as in the Numero- 
scope, characters are produced on a cathode-ray tube screen by 
means of combinations of signals applied to the deflection plates. 
A technique based on magnetic recording principles is used to 
generate the signals. 



408 PHYSICAL COMPONENTS AND METHODS 

The Monoscope. A piece of equipment commonly used in 
television transmitting stations is known as a monoscope. 2 *' 29 
This device is a small Iconoscope television transmitting tube 
with a picture or a diagram fixed on the inside of the envelope. 
A plate with a fairly high secondary-emission ratio, such as 
aluminum with its normal coat of oxide, is used in place of the 
mosaic. A figure is painted on the plate with a material, such as 
carbon ink, having a contrasting secondary-emission ratio. As 
the plate is scanned with an electron beam, a secondary-emission- 
current pattern of the diagram is produced. A device similar to 
the Harvard Numeroscope has been contemplated using a bank 
of 10 monoscopes each having a decimal digit built in. All 
could be operated from common sweep generators as could the 
Kinescope tubes on which the numbers are made visible. The 
output register would serve to key the proper monoscope into 
a row of Kinescope tubes arranged to be photographed as in the 
Numeroscope. 

Other Digital Recorders. The use of the Selectron as a digital- 
to-analog converter was suggested on page 395. An image 
plotted on a Selectron used as a converter could be photographed 
in order to make a permanent record of computer results. In 
order to keep the trace on the screen for a time long enough to 
be photographed, it might be necessary to introduce the same 
set of values cyclically for a short period of time. 

Electrical pulse-operated counter wheels also may be photo- 
graphed to provide a permanent digital record. 

15-5-2. Transcription from Automatic Analog to Written 
Form. It is often desirable to record the output of an analog 
machine for future reference and study. The output of the most 
common analog devices is a voltage or a shaft rotation. Various 
ways of recording these outputs are discussed in this section. 

Electrical Output Recorders. An important type of electrical 
analog device is the d-c or a-c network analyzer. The output 
of a network analyzer is the current, power, or voltage at a 
number of points in the network, the number being dependent 
upon the type and size of the network under analysis. As an 
aid to reading and transcribing to paper a large number of volt- 
ages, a recording voltmeter, such as the Brown Recorder, may be 
employed. These devices contain servosystems which position 
a pen on a linear track perpendicular to the direction in which a 



DATA-CONVERSION EQUIPMENT 409 

long sheet of paper can be moved in discrete steps. 30 The zero- 
voltage position of the pen and the scale factor may be adjusted 
so that any voltage of sufficiently high level can be recorded. 
If voltages are to be measured at 50 different places in the net- 
work analyzed, a probe may be applied to each of these in turn 
and the voltages recorded along 50 different axes on the paper. 
If some parameter is changed, the paper can be rolled back and 
the entire process repeated. Each set of dots, when connected 
or otherwise identified, results in a plot of values along some 
route through the network. 

The output of the Westinghouse Transient Analyzer is a pair 
of cyclical voltages, displayed on the screen of a cathode-ray 
oscilloscope to represent the wave form of a transient quantity. 
One of the cyclic outputs is usually a saw-tooth voltage to 
represent a time base. Still photographs are taken of the output 
curves for permanent records. 

The output of an electronic differential analyzer is also one 
or more voltages which vary in time. A display of an output 
voltage along one axis of a cathode-ray screen can be photo- 
graphed with a shutterless camera to produce a continuous curve 
as a function of time. Similarly, output voltages may be 
recorded on an electromagnetic oscillograph. 

Mechanical Curve Tracers. The output of a mechanical 
differential analyzer usually consists of the rotation angle of 
two or more shafts. A plot of one output as a function of 
another (usually machine time) can be obtained by arranging a 
pair of drive systems so that one output drives a marking instru- 
ment, such as a pencil, in one direction and the second output 
drives either a paper or the pencil in a direction perpendicular to 
the first. Thus, the pencil traces values on the paper in cartesian 
coordinates. A convenient mechanical arrangement is to place 
the paper over a cylindrical drum and have one drive cause the 
drum to rotate around its axis. When polar plots are desired, 
the pencil is moved in and out from the origin, and the paper is 
caused to rotate about the origin. 

Since the output mechanism of the integrators used in a dif- 
ferential analyzer will slip if loaded heavily, a servosystem is 
usually required to drive the curve-plotting device. The 
requirements for the servosystem are low static and dynamic 
error and good stability, or freedom from overshooting. The 



410 PHYSICAL COMPONENTS AND METHODS 

marking device may be a pencil, a pen, or an electric scribe, used 
with paper which is marked when current is passed through it. 
A paper of this type is called Teledeltos paper. 

Counter-wheel Printers. It may be desirable to sample the 
output of an analog quantity occasionally and write its numerical 
value. The first step in this process is to convert the continuous 
quantity to digital form and then to print the digital quantity 
by one of the methods described in Sec. 15-5-1. In the first 
MIT differential analyzer, for example, the output was on a 
mechanical counter with a printing wheel on the dials. At 
discrete intervals, a paper was pressed against these continuously 
rotating wheels, and a slightly blurred print of the representative 
numbers was obtained. The top speed of the commercial con- 
tinuous-input, Veeder-Root counters with printing wheels is 
17 counts per second. In the later model of the MIT differential 
analyzer, a bank of rotary counter switches is used to operate an 
electric typewriter, and a record of shaft position is typed from 
time to time. 



15-6. Radix Converters 

It may be recalled from descriptions of various arithmetic 
elements that circuits are more simple for arithmetic operations in 
the binary system than in the decimal system. Furthermore, 
most of the digital computing elements which have been used to 
date are employed in an on-off, or binary, fashion. Binary 
numbers, however, have more than three times as many digits 
as decimal numbers and are much less convenient for human 
operators to deal with. Therefore, it is usually desirable to have 
operators enter decimal numbers into a machine. Some thought 
has been given to the use of octal numbers at the input of a 
machine because of the ease of converting them to binary form. 
In any case, unless the operator inserts binary numbers, a change 
in radix takes place at some level in a digital computer. 31 

In the Harvard Mark I and the ENIAC, decimal arithmetic is 
performed, and the only radix conversion is at the switches and 
flip-flops, which represent the decimal digits in a binary manner. 
In the Harvard Mark II, the UNIVAC, and other computers, 
each decimal digit is represented by a group of four binary digits. 
Such a system is known as a coded-decimal system. A method of 



DATA-CONVERSION EQUIPMENT 411 

conversion of decimal numbers to a binary-coded system is 
described on page 289. 

This section includes a discussion of three types of radix 
conversion used in present or contemplated computers. They 
are the decimal-to-binary and binary-to-decimal conversion, 
using binary equipment, and the coded-decimal conversion. 

15-6-1. Decimal-to -binary Conversion Using Binary-arith- 
metic Elements. In the conversion of decimal numbers to 
binary numbers, the process varies somewhat depending upon 
whether the numbers are whole numbers with no fractional part, 
are fractional numbers with no whole part, or are a combination 
of the two. However, the general idea of the process is illus- 
trated by a single case. 

To convert whole decimal numbers to binary form using binary 
equipment, one might first find the binary equivalent of each 
digit. This process can be performed in a matrix similar to the 
one illustrated in Fig. 4-36 on page 43 and described as an octal- 
to-binary converter. Each digit might then be multiplied by 
the binary notation for 10 a number of times corresponding to 
the power of 10 with which it was associated in the original 
decimal number, and all the products summed. For example, 
to convert 132 to binary form, the digits are first converted to 
0,001, 0,011, and 0,010, respectively. The first of these is next 
multiplied twice by binary 10, or 1,010, the second once, and the 
third not at all. The resulting numbers are 1,100,100, 11,110, 
and 0,010, which are added to produce 10,000,100, which is the 
binary expression for 132. 

Since the only two arithmetic steps required to accomplish this 
conversion are multiplication by 1,010 and addition, a complete 
arithmetic unit is not actually required to perform this work. 
Therefore, either the automatic conversion of input data can be 
accomplished in the arithmetic unit of the machine at the expense 
of some computing time or it can be accomplished in a pre- 
liminary conversion unit at the expense of some extra equipment. 
In a modern computer, the time lost in the first case usually will 
be negligible. 

If the conversion is accomplished in a separate unit previous 
to the time when the problem is to be solved, there is a possi- 
bility of checking input data after conversion, thereby checking 
the conversion process as well as the operator. 



412 PHYSICAL COMPONENTS AND METHODS 

15-6-2. Binary-to-decimal Conversion Using Binary-arith- 
metic Elements. To convert a whole binary number to the 
decimal system, it is necessary only to divide successively by 
1,010, the binary equivalent of 10, and to convert the remainders 
to decimal digits in a matrix which is the inverse of the one 
required to convert from decimal to binary form. As an illus- 
tration, consider the transformation of the binary number 
10,000,100 into the decimal system. The form of the division 
required is as shown : 

1101 



)|10000100 
1010 
1101 
1010 



1100 
1010 
0010 Remainder 

1 



1010|1101 
1010 



0011 Remainder 




ioioJi 




0001 Remainder 

The three remainders are arranged in inverse order and converted 
to decimal notation with another matrix circuit similar to the 
one illustrated in Fig. 4-3a on page 42. Thus 0,001, 0,011, 
0,010 becomes 132, which may be recognized as the original 
decimal number. 

Again, only a limited amount of arithmetic equipment is 
required to perform the conversion, so there is a choice to make 
between programming the conversion through the main arith- 
metic unit or providing a converter in the output of the computer. 
Since the problem is again one of equipment economy versus 
economy of time, the method of solution chosen will depend on 



DATA-CONVERSION EQUIPMENT 



413 



both the type of problem which the computer is expected to 
solve and on the computer development budget. 

15-6-3. Coded-decimal Converters. When a coded-decimal 
system is used, the conversion process is shorter by one step than 
a binary-decimal conversion. Since each decimal digit is 

B + 




X 



X 



X 



X 



\-\ 



VX 



x 



x. 



x 



x~x 



x~x 



x 



\~x 



X 



X 



\-x. 



X 



X. 



\ 



X 



X 



X 



ON OFF ON OFF ON OFF ON OFF 

ORDER DIGIT 4 3 2 1 

Fig. 15-5. Diode matrix for converting from decimal to excess-3 coded 
decimal notation. 



represented by a four-digit binary number, the conversion may 
be made on a simple matrix arranged to give the coded-decimal 
equivalent of each digit in a decimal number. A converter of 
this type is illustrated in Fig. 15-5. 

The switch at the left-hand side of Fig. 15-5 is set on the 
decimal digit 5. According to the diode current-flow convention 
adopted in this text, this position of the input switch causes the 
positive voltage supply to be grounded through the resistor 
labeled 1, of the left-hand pair at the top of the diagram. This 



414 PHYSICAL COMPONENTS AND METHODS 

signifies that the fourth-order binary digit of the number into 
which 5 is to be converted is 1. Of the second pair of resistors at 
the top of the diagram, representing the third-order binary digit, 
the one labeled is grounded. Hence, the third-order digit is 0. 
Similarly, in both remaining pairs the side is grounded. The 
binary representation is therefore 1000. This is the binary 
representation for 8 and is therefore the proper representation 
for the decimal number 5 in an exccss-3 code. A study of the 
matrix shown in Fig. 15-5 will demonstrate to the reader that if 
any of the other input switch positions is selected a different 
binary number will be represented and each of these binary num- 
bers will be the proper representation of the number on which the 
10-position switch is set, plus 3. 

It may be of interest to compare this matrix with those shown 
in Figs. 4-3a and 4-36 on pages 42 and 43. 

There are other systems for coding decimal numbers in binary 
notation. The excess-3 code has been used as an example of one 
of these methods. It is the method of number representation 
used in the UNIVAC. 

15-7. Speed-changing Devices 

From a consideration of the numerous forms of data which are 
suitable for computations, it is evident that data may be gen- 
erated at vastly different speeds. Likewise, in a computer, 
certain operations can be performed faster than others. The 
most efficient machine design for the handling of a given problem 
is one which allows a continuous flow of data from beginning to 
end. This may require parallel operation of some of the slower 
units so that data may be treated as rapidly as in a faster unit. 
It may also require the use of intermediate storage mediums on 
which to store the data until time for the next operation. Data 
may be accumulated on some of these mediums at one rate of 
speed and released at another. 32-34 

A situation in which speed changing is obviously needed is 
the entry of manually prepared data into a high-speed machine. 
If several operators can prepare storage mediums for relatively 
high-speed automatic entry of data into the computer, a great 
deal of machine utilization time can be saved. Likewise, if the 
computation results are temporarily stored on an intermediate 
medium which can be removed from the machine, the data may 



DATA-CONVERSION EQUIPMENT 415 

be transcribed manually or automatically at a suitable rate 
without interfering with machine operation on a new problem. 
Several transcribers can then operate on the mediums simul- 
taneously, thus providing compensation for the difference in 
speed between the output of data and printing. Also, use of 
intermediate storage mediums allows problem preparation at 
various locations; the problems may then be sent to a central 
location for actual computation. 

15-7-1. Punched Cards and Punched Tape. Cards and tape 
may be punched by several operators simultaneously and then 
fed into an automatic device at top automatic speed. Their 
top speed, however, is somewhat low in comparison with the 
operating speed of an electronic computer. For example, 
punched cards are fed into the ENIAC only at the rate of one 
every half second, but arithmetic operations can be performed on 
the data in a few milliseconds. In problems which require a 
large number of arithmetic steps after each unit of data is 
inserted, this mismatch in speed might not be a great handicap. 
(As indicated in Sec. 10-3-1, the ENIAC, as currently used, is 
not limited by its card-handling speed.) 

In the Bell Telephone Laboratories machines which use tele- 
type tape as the input medium, a seven-digit decimal number, 
or command, may be entered into the machine in 0.2 second, 
added in 0.3 second, and multiplied in 1.0 second. It has been 
found from experience that numbers usually reach the machine 
at about the proper rate, but the input of commands is somewhat 
slow. This indicates roughly that input time should be about 
one-fifth of a multiplication time. 

15-7-2. Magnetic Mediums. Several characteristics of mag- 
netic mediums make them desirable for use with speed changers. 
Data may be recorded on them at low speeds and removed at 
fairly high speeds, and they can be erased and used over and over 
again. The original cost and volume per digit is relatively low, 
and it is easy to correct errors made by operators in preparing 
magnetic mediums. One disadvantage in their use is the extra 
equipment required to read from magnetic mediums when they 
are moved at low speeds. 

It is estimated that a 10-digit decimal number could be entered 
into a machine from magnetic mediums in 2 milliseconds and 
that numbers could be recorded on the tape at the same rate. 



416 PHYSICAL COMPONENTS AND METHODS 

15-7-3. Photographic Film. Like magnetic mediums, photo- 
graphic film is another medium which can be recorded on at 
either a fairly high speed or a low speed and can be read over a 
similar range of speeds. Also, it is commercially available in 
several widths; driving equipment is well developed; and data 
can be stored compactly. Since it can be used only once and 
since it is fairly expensive to supply and process, its use seems to 
be most suitable in places where it will be referred to repeatedly. 
Thus, film might be used to provide rapid access to tables of 
functions. Film is also interesting as an output medium because 
decimal numbers written on it at high speeds can be printed on 
photographic paper in the form of tables of output data. It is 
also versatile in that continuous curves as well as numbers may 
be recorded. 

REFERENCES 
Mechanical Counters 

1. Bush, V., A New Type of Differential Analyzer, Journal of the Franklin 
Institute, Vol. 240, No. 4, pp. 292-295 (October, 1945). 

2. Murray, F. J., The Theory of Mathematical Machines (King's Crown 
Press, New York, 1948), pp. I-1-I-9. 

Film-type Converters 

3. Jones, F. D., Ingenious Mechanisms for Designers and Inventors, (Indus- 
trial Press, New York, 1936), Vol. I, pp. 71-77, Vol. II, pp. 164-188. 

Pulse Code Modulation 

4. Black, H. S., and J. O. Edson, Pulse Code Modulation, Transactions 
of the American Institute of Electrical Engineers, Vol. 66, pp. 895-899 
(1947). 

5. Carbrey, R. L., Decoding in P. CM., Bell Laboratories Record, Vol. 
XXVI, No. 11, pp. 451-455 (November, 1948). 

6. Goodall, W. M., Telephony by Pulse Code Modulation, Bell System 
Technical Journal, Vol. XXVI, No. 3, pp. 395-409 (July, 1947). 

7. Greig, D. D., Pulse-Count Modulation, Electrical Communications, 
Vol. 24, No. 3, pp. 287-296 (September, 1946). 

8. Meachan, L. A., and E. Peterson, An Experimental Multi-Channel 
Pulse Code Modulation System of Toll Quality, Bell System Technical 
Journal, Vol. XXVII, pp. 1-43 (January, 1948). 

9. Oliver, B. M., J. R. Pierce, and C. E. Shannon, The Philosophy of Pulse 
Code Modulation, Proceedings of the Institute of Radio Engineers, Vol. 
36, No. 11, pp. 1324-1331 (November, 1948). 

10. Sears, R. W., Electric Beam Deflection Tube for Pulse Code Modula- 
tion, Bell System Technical Journal, Vol. XXVII, pp. 44-57 (January, 
1948). 



DATA-CONVERSION EQUIPMENT 417 

11. Unsigned, Coded Pulse Modulation Minimizes Noise, Electronics, 
Vol. 20, No. 12, pp. 126-131 (December, 1947). 

Rotation Converters 

12. Brown, G. S., and D. P. Campbell, Principles of Servomechanisms 
(John Wiley & Sons, Inc., New York, 1948). 

13. Bush, V., A New Type of Differential Analyzer, Journal of the Franklin 
Institute, Vol. 240, No. 4, pp. 292-295 (October, 1945). 

14. Hall, A. C, Applications of Circuit Theory to the Design of Servo- 
mechanisms, Journal of the Franklin Institute, Vol. 242, No. 2, pp. 
279-307 (October, 1946). 

15. Sobczyk, Andrew, Stabilization of Carrier Frequency Servomechanisms, 
Journal of the Franklin Institute, Vol. 246, No. 1, Nos. 2 and 3; Part I: 
pp. 21-43; Part II: pp. 95-121; Part III: pp. 187-213 (July, August, 
September, 1948). 

Conversion of Data from Written to Automatic Digital Form 

16. Bush, G. L., Photoelectric Transmitting Typewriter, Electronics, Vol. 
20, No. 4, pp. 96-99 (April, 1947). 

17. Dudley, H., Remaking Speech, Journal of the Acoustical Society of 
America, Vol. 11, p. 169 (May, 1939). 

18. Dvorak, A., There Is a Better Typewriter Keyboard, National Business 
Education Quarterly, Vol. XII, No. 2, pp. 51-58 and 66 (December, 
1943). 

19. Hasley, R. J., and J. Swaffield, Analysis Synthesis Telephony with 
Special Reference to the Vocoder, Journal of the Institution of Electrical 
Engineers, Vol. 95, Part III, pp. 391-410 (September, 1948). 

20. Watson, E. F., Fundamentals of Teletypewriters Used in the Bell 
System, Bell System Technical Journal, Vol. 17, pp. 620-639 (October, 
1938). 

Conversion of Data from Written to Automatic Analog Form 

21. Greenwood, Holdam, and MacRae, Electronic Instruments, Radiation 
Laboratory Series (McGraw-Hill Book Company, Inc., New York, 1948), 
pp. 362-372. 

22. Hazen, H. L., J. J. Jaeger, and G. S. Brown, Automatic Curve Follower, 
Review of Scientific Instruments, Vol. 7, pp. 353-357 (September, 1936). 

23. Walther, A., Development of Mathematical Instruments in Germany 
1939-1945 (translated by British Naval Gunnery Mission). Pamphlet. 

Transcription from Automatic Digital to Written Numerical Form 

24. Burnett, C. E., The Monoscope, RCA Review, Vol. 2, pp. 414-420 
(April, 1938). 

25. Fuller, Harrison W., The Numeroscope, Annals of the Computation 
Laboratory of Harvard University, Vol. XVI, pp. 238-247 (1948). 

26. Fuller, H. W., Numeroscope for Cathode Ray Printing, Electronics, 
Vol. 21, No. 2, pp. 98-102 (February, 1948). 



418 PHYSICAL COMPONENTS AND METHODS 

27. Mansbcrg, II. P., A New, Versatile Camera for the Cathode Ray Oscillo- 
graph, Oscillographer, Vol. 10, No. 4, pp. 6-8 (October-December, 1948). 

28. Sparks, S., and R. G. Kreer, Tape Relay System for Radio-Telegraph 
Operation, RCA Review, Vol. 8, No. 3, pp. 393-426 (September, 1947). 

29. Zworykin, V. K., and G. A. Morton, Television (John Wiley & Sons, 
Inc., New York, 1940), p. 327. 

Transcription from Automatic Analog to Written Form 

30. Kane, J. L., Conversion of Brown Recorder, unpublished report (Douglas 
Aircraft Co., Inc., El Segundo, Calif., September, 1947). 

Radix Converters 

31. Mauchly, John W., Theory and Techniques for the Design of Electronic 
Digital Computers, Vol. Ill (Moore School of Electrical Engineering, 
University of Pennsylvania, Philadelphia, June, 1948), pp. 25-1-25-8. 

Speed-changing Devices 

32. Alexander, S. N., Input and Output Devices for Electronic Digital 
Computing Machinery, Annals of the Computation Laboratory of Harvard 
University, Vol. XVI, pp. 248-253 (1948). 

33. Alt, F. L., A Bell Telephone Laboratories Computing Machine, Mathe- 
matical Tables and Other Aids to Computation, Vol. Ill, No. 21, pp. 1-13 
(January, 1948). 

34. Sheppard, C. Bradford, Transfer between External and Internal 
Memory, Annals of the Computation Laboratory of Harvard University, 
Vol. XVI, pp. 267-273 (1948). 



CHAPTER 16 

SPECIAL TECHNIQUES AND EQUIPMENT FOR POSSIBLE 
USE IN COMPUTING SYSTEMS 

16-1. Introduction 

As problems of greater complexity are presented for solution 
on high-speed computers, the question arises as to what speeds 
and improvements in computing techniques can be expected in 
the future. To a large extent the answer to this question depends 
upon how well the components and techniques of other fields 
of endeavor can be utilized in computing systems. Although 
it is difficult to predict the exact trend of future development in 
automatic computing, it seems clear that higher pulse rates are 
desirable and that greater simplicity, reliability, and tube 
economy will be sought. Attempts are also being made to find 
dependable elementary potential digital computing components 
and storage mediums which have more than two stable states 
so that arithmetic operations with radices higher than 2 can be 
performed in a straightforward manner. Success in these 
attempts will increase the capacity of storage mediums and may 
possibly increase the speed of computation. 

16-2. New Techniques 

Digits are stored in sonic delay lines as pulses of a high- 
frequency carrier voltage. Likewise, if low-velocity waveguides 
are used as a storage medium, an ultrahigh-frequency carrier 
voltage is required. It may be convenient to transfer digits 
through other sections of a computer as high-frequency pulses. 
This practice is followed in television and in radar receivers, 
where pulses are partially amplified in intermediate-frequency 
amplifiers rather than totally in video amplifiers, since the inter- 
stage coupling problem is less complicated for an amplifier of 
30 + 5 megacycles than for an amplifier with half the band 
pass yet requiring d-c amplification. The cost advantage of 
high-frequency components in comparison with video circuit 

419 



420 PHYSICAL COMPONENTS AND METHODS 

components may be increased when printed-circuit techniques 
are developed further. Also, high-frequency circuits are more 
convenient when band-pass characteristics are desired. 

The field of wide-band-amplifier design has recently had some 
special developments, of which the distributed amplifier is an 
example. This type of amplifier has many operational features 
in common with the traveling-wave tube, which is discussed 
below. However, it is different in that it uses standard com- 
ponents. When amplifier tubes are cascaded in the conventional 
manner, the shunt capacities of parallel tubes increase in the 
same ratio as the gain. Therefore, it is not possible to realize a 
gain-bandwidth product which is any greater than that of a 
single tube. In other words, as more tubes are added, gain is 
increased at the expense of bandwidth. In distributed amplifiers, 
the tubes are separated by inductances, and the resulting system 
simulates a transmission line. As a result, the gains of the tubes 
are increased and the distributed capacities are not. Therefore, 
the bandwidth of the whole amplifier can approach the figure of 
merit, or unit gain bandwidth, of a single tube. 

A development-type amplifier employing 14 tubes has a gain 
of 18 decibels over the frequency band of to 200 megacycles. 
An extension of this technique may be applicable to computing 
circuits. 

The degree of compactness with which large numbers can be 
represented is a function of the radix of the number system being 
used. The largest number n which can be represented with a 
digits of radix r is r° — 1. Since a given number can be repre- 
sented by a smaller number of digits in a larger radix system, 
there is a resulting saving in time when the higher radix number 
is stored in serial storage, and a resulting saving in number of 
channels when the number is stored in parallel channels. A 
saving of 58 per cent can be gained in going from a binary to a 
ternary system. A smaller percentage gain is realized in going 
from a radix 3 to a radix 4 system. 

One means of increasing the radix of computing components is 
to represent individual digits by corresponding frequencies. For 
engineering convenience, those frequencies might be chosen 
which could be stored directly in the storage medium. This 
system is also interesting because some of the elementary arith- 
metic operations may be accomplished by beating frequencies 



SPECIAL TECHNIQUES AND EQUIPMENT 421 

together. It also has the advantage that components which 
are essentially video amplifiers in on-off pulse techniques may 
be replaced with elements which are similar to intermediate- 
or high-frequency amplifiers. A certain amount of back- 
ground development has already been done in frequency-shift 
telegraphy. 1,3 ' 4 

It is interesting to note that the increase in access time and 
storage capacity of a medium by use of higher radix frequency- 
shift methods is at the expense of greater bandwidth in the storage 
medium and the computing circuits. Recent developments 
indicate that it may be easier to develop circuits with greater 
bandwidth than to increase the cycling rate of an EPDCC such 
as a flip-flop circuit. 

16-3. Special Electronic Tubes and Crystals 

The electronic elements described in this section are of interest 
as high-frequency amplifiers and as means for increasing the 
simplicity, economy, and dependability of circuits. 

16-3-1. The Traveling-wave Tube. An electronic tube has 
recently been developed at the Bell Telephone Laboratories 6,7 
which provides amplification over a very wide band of fre- 
quencies. The amplification results from the interaction 
between an electron beam and a slowly moving electromagnetic 
field. The field is produced by a helical transmission line; the 
phase velocity of the field is reduced to about one-thirteenth 
the velocity of light. The property of helical delay lines to 
reduce phase velocity was discussed on page 351. The electron 
beam in the tube is passed along the axis of the helix at a velocity 
slightly higher than that of the field, and energy is transferred 
from the electron stream to the traveling field. In an early 
model 5 an input signal was amplified 23 decibels as it traversed 
the length of the helix. The helical line replaces the resonant 
circuits in conventional high-frequency amplifiers, and, as a 
result, the bandwidth for a 3-decibel decrease in amplification is 
800 megacycles at a mean frequency of 3,000 megacycles. 

16-3-2. Klystrons. Another high-frequency tube, the kly- 
stron, has been used extensively as a local oscillator in microwave 
radio and radar receivers. 8 ' 9,11 ' 12 In the klystron an electro- 
magnetic field is produced in a resonant cavity; the velocity of 
an electron beam traversing the cavity is modulated by bunching 



422 PHYSICAL COMPONENTS AND METHODS 

electrodes so that the beam of electrons is bunched to form 
current pulses. The dimensions are of the order of a wavelength 
at the frequencies handled, so that conventional klystrons 
become inconveniently large at frequencies much below 1,000 
megacycles. Klystrons may be used as voltage amplifiers, 
power amplifiers, and oscillators with relatively high efficiency 
and power-handling capacity. Some recent work on klystrons 
used as generators of very short pulses has resulted in pulses of 
adjustable duration as low as 0.2 millimicrosecond. 10 The 
experiments were performed at a repetition rate of 210 megacycles 
with a klystron in which the resonant cavity had been replaced 
with a nonresonant line. With present-day circuits and produc- 
tion methods it is difficult to utilize pulses of duration shorter 
than 0.1 microsecond in computers, but the full capabilities of 
the klystron pulse generator may become useful in computers 
as pulse-handling techniques are improved. 

16-3-3. Miniaturized Tubes. Counterparts for most of the 
common radio receiving tubes are now manufactured in miniature 
form. 14 These miniature tubes occupy about three-tenths the 
volume of conventional tubes of similar type and make possible 
a considerable reduction in size and weight of electronic equip- 
ment. Miniature tubes seem to be dependable and as well 
suited to computer systems as the larger tubes. 

Another group of tubes called subminiatures has been devel- 
oped. The diameter of these tubes is approximately that of 
an ordinary lead pencil, and flexible leads are used rather than 
base pins. The tubes are available in both the filament and the 
cathode types. A typical application is found in the modern 
hearing aid. 

Manufacturers' catalogs of miniature tubes are available, 15 
and a complete list of subminiature tubes has been compiled 
by the Signal Corps. 13 Miniature tubes are manufactured by 
RCA, Raytheon, and Sylvania; subminiature tubes are manu- 
factured by Raytheon, Sylvania, and Sonotone. 

16-3-4. Transistor. A new type of amplifier called the tran- 
sistor was recently announced by the Bell Telephone Labora- 
tories. 16-18 It seems likely that this device will simplify com- 
puter circuits considerably. The transistor is similar to the 
crystal rectifier which was described in Chap. 3 and consists of a 
block of germanium crystal with which two cat whiskers of 



SPECIAL TECHNIQUES AND EQUIPMENT 423 

2-mil tungsten wire are in contact. These two wires are spaced 
2 mils apart and form one side of the input and the output 
circuits, respectively. The germanium crystal is at the common 
connection between input and output. A gain of 20 decibels 
at 25 per cent efficiency was measured on the model announced. 
The transistor has an input impedance which is only 1 per cent of 
its output impedance, so that interstage coupling must necessarily 
differ from that used in vacuum-tube circuits. Transit time 
establishes a top frequency limit of about 10 megacycles for the 
transistor. 

The advantages of this element are its small size, low power 
drain, sturdy construction, long life, and cool operation. It will 
probably be competitive with the electron tube in total cost per 
stage. 

16-4. Amplitude- and Frequency-discrimination Circuits 

The action of the gates, flip-flops, and relays which have 
been discussed previously depends upon the ability of these units 
to discriminate between low-level spurious voltages and high- 
level signals. 22 The units heretofore described discriminate on 
the basis of voltage amplitude. Equipment is also available 
which discriminates on the basis of frequency. 

The possibility of representing digits by discrete frequencies 
is introduced in the second section of this chapter. In such a 
system, it may be desirable to convert the carrier frequencies 
into corresponding d-c voltage levels for arithmetic operations 
such as those described in Sec. 13-4-4 on amplitude adders. 
Conversion from frequency to voltage amplitude is accomplished 
by means of a frequency discriminator. 

Frequency discriminators were first used in conjunction with 
automatic frequency controls and later in radio receivers using 
frequency modulation. There are many competing methods of 
frequency discrimination. 19_21,26 The better known of these 
are the Seeley-Foster and the Armstrong methods. Recent 
variations of the Seeley-Foster circuit are the ratio detectors, 23,27 
and a variable bandwidth discriminator developed at MIT. 28 

A discriminator which is used at microwave frequencies is 
described by Pound. 24 - 25 This discriminator makes use of two 
waveguide arrangements known as magic T's and a cavity which 
is resonant at the mid-frequency of the frequency-modulation 



424 PHYSICAL COMPONENTS AND METHODS 

signal. The cavity is not loaded, so that it may have a high Q. 
The output of this discriminator is the difference in rectified 
voltage at two points in the waveguide. The relative phase of 
voltage received at these two points is such that the voltage 
difference is a function of the applied frequency. 

16-5. Frequency-stabilizing Systems 

Computers using cyclic storage mediums require pulse-repeti- 
tion-rate stabilization in order to preserve synchronization. 
Accurately held high frequencies may be used to time lower 
frequency pulse generators. Several circuits have been developed 
which perform the function of maintaining microwave oscillators 
at nearly constant frequency. One circuit, developed by Pound 
at the MIT Radiation Laboratory, 31 - 32 uses the microwave dis- 
criminator described in the previous section as a frequency- 
sensing device. The output of this discriminator is amplified 
and applied to an electrode of the oscillator through a negative- 
feedback loop, and the oscillator frequency is held near the 
frequency of the resonant cavity in the discriminator. 

Another such device, developed at RCA, 30 employs the 
spectral line of a gas as the frequency reference with which the 
oscillator frequency is compared. The voltage output of a 
comparison circuit is used in a negative-feedback, or servo, 
loop to control the frequency of the oscillator. Multivibrator 
step-down circuits can be used to obtain lower frequencies from 
these high-frequency standards. 29 

16-6. Unconventional Circuit Fabrication 

Interest in subminiature electronic equipment stems from its 
extensive and successful use in certain war weapons. One of 
these weapons, the VT fuse, was made in mass production using 
techniques specially developed for rapid fabrication. Mass- 
production methods known commonly now as printed-cir- 
cuit techniques were used in making fuses for trench-mortar 
shells. 34 Interest in these methods has been displayed by the 
electronic manufacturing industry, and subsequently exten- 
sive development work to expand and improve these techniques 
was undertaken. 

16-6-1. Methods of Printing Circuits. Brunetti and Curtis 34 
have listed six general methods for the mass production of printed 
circuits: 



SPECIAL TECHNIQUES AND EQUIPMENT 425 

1. Painting. A special conducting paint is applied to a 
prepared ceramic surface placed under a stencil. It is 
possible to wire the circuit components as well as to create 
the components with conducting paint. For example, a 
certain deposited thickness, width, and length of a special 
resistor paint may be made to represent a given value of 
resistance, while a change of any of the dimensions or of 
the paint itself may be made to represent a value different 
from the first. It is possible to print inductances and some 
capacitors by an extension of this method. 

2. Spraying. This method is similar to the painting technique 
discussed above with the exception that the coatings are 
sprayed on the ceramic plate placed under the circuit 
stencil. 

3. Chemical deposition. It is possible to construct an elec- 
tronic circuit by electrolysis. The wiring is in the form 
of a metallic deposit on the special ceramic plate. 

4. Vacuum process. Less used but interesting techniques for 
printing circuits are those of cathode sputtering and 
metallic evaporation. 

5. Die-stamping. This method allows the interconnecting 
wires of the circuit and some high-frequency inductances 
to be die-punched out of a solid sheet. 

6. Dusting. The metallic conductors may be dusted with a 
small brush onto a ceramic plate under a stencil and then 
fused on the surface. 

16-6-2. Examples of Commercially Available Units. A typical 
example of the commercial application of the printed circuit is to 
be found in a subminiature audit) amplifier called the Ampec. 
This unit is available in production quantities from the Centralab 
Division of the Globe Union Co., Inc., Milwaukee, Wisconsin. 
The Ampec is a three-tube amplifier consisting of two voltage 
amplifiers followed by a power amplifier. The unit does not 
include the volume control, output transformer, or battery 
power unit; it is fabricated with printed-circuit techniques entirely 
and measures 2.250 inches long, 1.15 inches wide, and weighs 0.63 
ounce. This unit has a voltage gain of 4,000 when used with a 
30-volt plate battery and a 1.5-volt filament cell. The plate- 
battery drain is approximately 1 milliampere and the filament 
drain is 40 milliamperes. Although the unit is somewhat micro- 



426 PHYSICAL COMPONENTS AND METHODS 

phonic, it is used extensively as the basic unit for several 
commercial hearing aids. 

The manufacturers of the Ampec have also made available 
printed-circuit-component combinations, such as interstage 
coupling units, that save space in the more conventional radio 
applications. 

Another printed-circuit technique which has been greatly 
emphasized is the die-stamping process. The Franklin Corpora- 
tion manufactures an inductance loop stamped out of sheet 
copper which serves as a high-Q tuned element of the table- 
model radio receivers. 

16-6-3. Features of Printed Circuits. The major features of 
printed-circuit techniques are (1) size reduction, (2) uniformity, 
and (3) economy for large numbers of units. 

Size reduction has been made possible by printed circuits 
to the extent that the volume occupied by the wiring and sup- 
porting plate is only a small percentage of the total volume of 
the unit. The development of truly subminiature units is 
dependent not upon further size reduction in circuitry but 
rather upon the development of smaller inductors, transducers, 
and power supplies. 

The application of the printed-circuit technique to the fabrica- 
tion of electronic equipment has placed the manufacture of 
this equipment in the true mass-production class, where machine 
operations can be accurately controlled and the human errors 
generally encountered in wiring work can be minimized. 

A great advantage in the use of printed circuits is the con- 
siderable reduction in man-hours required to fabricate the units. 
A further saving is made in the man-hours of inspection time 
required to check the work. Obviously, the printed-circuit 
technique requires specialized and complicated manufacturing 
equipment to fabricate printed-circuit assemblies. The expense 
involved for such equipment is not justified for small quantities; 
large production schedules are necessary before the techniques 
can be profitably applied. 

16-6-4. Plug-in and Potted Circuits. The fabrication of 
printed-circuit subassemblies suggests the use of plug-in units 
for these assemblies. Tests made at the Bureau of Standards 
showed that it was possible to pot printed circuits in a special 
casting resin in such fashion as to permit the plugging in of the 



SPECIAL TECHNIQUES AND EQUIPMENT 427 

complete subassemblies. By the use of special resins, the fre- 
quency of operation of the printed circuit may be extended into 
the ultrahigh-frequency region without a noticeable decrease in 
efficiency owing to the enclosure in the casting resin. The use 
of casting resins, however, requires critically controlled condi- 
tions, and consequently mass production of potted plug-in units 
depends upon the development of complex process controls. 

REFERENCES 
Improved Techniques 

1. Darey, J. R., and A. L. Matte, Frequency Shift Telegraphy — Radio and 
Wire Applications, Bell System Technical Journal, Vol. 27, No. 2, pp. 
265-304 (April, 1948). 

2. Ginzton, Edward L., et al., Distributed Amplification, Proceedings of the 
Institute of Radio Engineers, Vol. 136, No. 8, pp. 956-969 (August, 1948). 

3. Hatfield, L. E., Frequency Shift Radio Transmission, Proceedings of the 
Institute of Radio Engineers, Vol. 36, No. 1, pp. 116-120 (January, 1948). 

4. Vanderlippe, R. A., Frequency Shift Radio Teletype in World War II, 
Bell Laboratories Record, Vol. XXV, No. 4, pp. 442-446 (December, 1947). 

Traveling-wave Tubes 

5. Kompfner, Rundolf, The Traveling Wave Tube as Amplifier at 
Microwaves, Proceedings of the Institute of Radio Engineers, Vol. 35, 
No. 2, pp. 124-127 (February, 1947). 

6. Pierce, J. R., and L. M. Field, Traveling Wave Tubes, Proceedings of 
the Institute of Radio Engineers, Vol. 35, No. 2, pp. 108-110 (February, 
1947). 

7. Pierce, J. R., Theory of the Beam Type Traveling Wave Tube, Proceed- 
ings of the Institute of Radio Engineers, Vol. 35, No. 2, pp. 111-123 
(February, 1947). 

Klystrons 

8. Hamilton, D. R., J. K. Knipp, and J. B. H. Kuper, Klystrons and 
Microwave Triodes, Radiation Laboratory Series, Vol. 7 (McGraw-Hill 
Book Company, Inc., New York, 1948). 

9. Harrison, A. E., Klystron Oscillators, Electronics, Vol. 17, No. 11, pp. 
100-107 (November, 1944). 

10. Hasted, J. B., Milli-Microsecond Pulse Generation by Electron Bunch- 
ing, Proceedings of the Physical Society, Vol. 60, p. 397 (Apr. 1, 1948). 

11. Pierce, J. R., and W. G. Shepherd, Reflex Oscillators, Bell System Techni- 
cal Journal, Vol. 26, No. 3, pp. 460-682 (June, 1947). 

12. Varian, R. H., and S. F. Varian, A High Frequency Oscillator and 
Amplifier, Journal of Applied Physics, Vol. 10, No. 5, pp. 321-327 
(May, 1939). 



428 PHYSICAL COMPONENTS AND METHODS 

Miniaturized Tubes 

13. Evans Signal Laboratory, Handbook of Subminiature Tubes (Signal Corps 
Engineering Laboratories, Bradley Beach, N.J., July, 1947). 

14. Green, N. H., Miniature Tubes in War and Peace, RCA Review, Vol. 8, 
No. 2, pp. 331-341 (June, 1947). 

15. Radio Corporation of America, RCA Quick Reference Chart Miniature 
Tubes (RCA Tube Department, Harrison, N.J.). 

Transistor 

16. Bardeen, J., and W. H. Brattain, The Transistor, A Semi-Conductor 
Triode, Physical Review, Vol. 74, No. 2, pp. 230-231 (July 15, 1948). 

17. Becker, J. A., Transistor Characteristics, Electronics, Vol. 22, No. 1, p. 
132 (January, 1949). 

18. Fink, D. G., and F. H. Rockett, The Transistor— A Crystal Triode, 
Electronics, Vol. 21, No. 9, pp. 68-71 (September, 1948). 

Amplitude- and Frequency-discrimination Circuits 

19. Barco, A. A., C. M. Kimball, and S. W. Seeley, Generation and Detec- 
tion of FM Waves, RCA Review, Vol. VI, No. 3, pp. 269-286 (January, 
1942). 

20. Crosby, M. G., Communication by Phase Modulation, Proceedings of 
the Institute of Radio Engineers, Vol. 27, No. 2, pp. 126-131 (Feburary, 
1939). 

21. Foster, D. C, and S. W. Seeley, Automatic Tuning, Simplified Circuits, 
and Design Practice, Proceedings of the Institute of Radio Engineers, 
Vol. 25, No. 3, p. 289 (March, 1937). 

22. Lewis, W. B., Electrical Counting (The Macmillan Company, New 
York, 1943), pp. 71-76. 

23. Maurice, D., and R. J. H. Slaughter, F. M. Comparison Test between 
Phase Discriminator and Ratio Detectors, Wireless World, Vol. 54, pp. 
103-106 (March, 1948). 

24. Pound, R. V., Microwave Mixers, Radiation Laboratory Series, Vol. 16 
(McGraw-Hill Book Company, Inc., New York, 1948), pp. 190-231. 

25. Roder, Hans, Theory of the Discriminator Circuit for Automatic 
Frequency Control, Proceedings of the Institute of Radio Engineers, Vol. 
29, No. 5, pp. 590-596 (May, 1938). 

26. Seeley, S. W., Balanced Phase Shift Discriminators, RCA Licencee 
Bulletin LB-611 (Mar. 15, 1946), RCA Licencee Bulletin LB-645 (Sept. 
15, 1945). 

27. Terman, F. E., Radio Engineers' Handbook (McGraw-Hill Book Com- 
pany, Inc., New York, 1943), pp. 585-588. 

28. Tuller, W. G., and T. P. Cheatham, Jr., Adjustable Band with FM 
Discriminator, Electronics, Vol. 20, No. 9, pp. 117-119 (September, 
1947). 

Frequency-stabilizing Circuits 

29. Davis, K. H., Multivibrator Stepdown by Fractional Ratios, Bell 
Laboratories Record, Vol. XXVI, No. 3, pp. 114-118 (March, 1948). 



SPECIAL TECHNIQUES AND EQUIPMENT 429 

30. Hershberger, W. D., and L. W. Norton, Frequency Stabilization with 
Microwave Spectral Lines, RCA Review, Vol. 9, No. 1, pp. 38-49 (March, 
1948). 

31. Pound, R. V., Frequency Stabilization of Microwave Oscillators, Pro- 
ceedings of the Institute of Radio Engineers, Vol. 35, No. 12, pp. 1405-1415 
(December, 1947). 

32. Tuller, W. G., Recent Developments in Frequency Stabilization of 
Microwave Oscillators, Proceedings of the Institute of Radio Engineers, 
Vol. 36, No. 6, pp. 794-800 (June, 1948). 

Unconventional Circuit Fabrication 

33. Brunetti, C, and W. S. Hinman, Radio Proximity Fuse Design, Journal 
of Research of the National Bureau of Standards, RP 1723, Vol. 37, pp. 
1-13 (July, 1946). 

34. Brunetti, Cledo, and Roger W. Curtis, Printed Circuit Techniques, 
Proceedings of the Institute of Radio Engineers, Vol. 36, No. 1, pp. 121- 
165 (January, 1948). 

35. Greenwood, Holdam, and MacRae, Electronic Instruments, Radiation 
Laboratory Series, Vol. 21 (McGraw-Hill Book Company, Inc., New 
York, 1948), pp. 700-704. 



CHAPTER 17 
FACTORS AFFECTING CHOICE OF EQUIPMENT 

17-1. Introduction 

This chapter is concerned with the means by which computing 
equipment may be made accurate and reliable; there is also 
presented here a summary of the capacity and cost of storage 
mediums and an approximation of the storage, reading, and writ- 
ing times to be expected. Although there are other factors 
affecting the choice of equipment, these cannot be summarized 
readily and are implicit in the foregoing discussion of equipment 
characteristics. Furthermore, the high-speed computing art 
is not in a mature enough state to allow an extensive comparison 
of equipment which has been thoroughly tested. However, 
the following material should indicate the orders of magnitude 
involved, and the degree of complexity required to meet given 
equipment specifications. 

17-2. Reliability of Components 

The operating time of a large-scale computer is very valuable; 
occasional failures, therefore, are expensive as well as annoying. 
There are at least two approaches to the problem of reducing 
equipment failure during operation. One method is to make 
each component so reliable that it will last longer than the period 
decided upon as the useful life of the equipment. Although it 
is difficult to guarantee absolute reliability, if the statistical life 
expectancy of the component is projected far enough into the 
future, the actual rate of failure should prove fairly low during, 
the early years of machine operation. As a second approach, 
means may be taken to extend as far as possible the operating 
period of equipment considered to have a limited life expectancy 
and to detect incipient failures before they occur. 

It is not enough to life-test one or two models of a component 
under normal operating conditions to determine whether or not 
it is suitable for use in a computer. It is necessary as well to 

430 



FACTORS AFFECTING CHOICE OF EQUIPMENT 431 

ascertain that when manufactured parts are used together, any 
part having a value within the manufacturing tolerance range 
will work reliably with any combination of other components 
which are also within the prescribed range. Since it is imprac- 
tical to test all possible combinations of values during design, 
circuits should operate with large factors of safety for voltages, 
pulse shapes, and repetition rates so that reasonable variations 
in components will not affect operation adversely. Furthermore, 
circuits should be designed so that failure of one unit does not 
overload other units and also cause their eventual failure. This 
condition may be difficult to carry out completely, but, at least, 
interdependent parts should be assembled in units so that all 
the parts which may be overloaded by the failure of one part are 
located together and can be replaced by exchanging units. Inter- 
changeable chassis are also desirable so that repairing and 
periodic testing of units can be accomplished simultaneously 
with normal operation. 

There are important factors in the philosophy of circuit 
design which also affect reliability. For example, time allow- 
ance between operations should be adequate to ensure correct 
sequences even when individual circuits lose their original 
adjustment. Allowances should be made for variations in signal 
voltages and in power-supply voltage. Direct coupling is often 
required between stages so that coupling capacitors will not 
become charged by large pulses and hold subsequent stages 
conducting longer than desired. Some means of detecting an 
instantaneous failure of the primary power source is required, 
since such a failure might otherwise cause an undetected error. 
The use of latch-type relays is desirable in a relay computer so 
that values will be held during the period between a failure and 
its repair. Finally, it should be recognized that many manu- 
facturers do not supply life-expectancy data on mechanical 
apparatus and that expensive tests may be required to get this 
information. 

In order to extend the life of individual components, thereby 
increasing operational reliability, it is necessary to follow recog-* 
nized design practices in each particular field. Mechanical 
devices must be designed for good wearing qualities and with 
adequate factors of safety in strength. Electrical apparatus, 
likewise, must be used at voltages and currents well under rated 



432 PHYSICAL COMPONENTS AND METHODS 

values. Cooling equipment must be provided wherever required 
in order to maintain safe temperatures in critical parts, and air 
conditioning is advisable for relay computers in order to minimize 
trouble from dust accumulations. 

In order to present a comprehensive view of the types of 
component failure which limit the reliability of large-scale 
computer operation, a number of specific examples of com- 
ponent failure are presented. Some of them have been taken 
from the operational history of the ENIAC. 

17-2-1. Vacuum Tubes and Crystals. It has been reported 
that in the ENIAC the most common source of trouble is in the 
18,000 vacuum tubes. Every tube in the machine undergoes a 
routine operational check of some sort each week. The practice, 
for example, is to test each of the several hundred three-tube 
amplifiers each week as a unit and to replace all three tubes if 
one of these amplifiers is not operating satisfactorily. During 
the first eleven months of 1949 about 2,000 tubes were replaced 
each month. About half of these were ac-tually bad at the time 
they were replaced. 

An attempt was made to increase the tube life in the ENIAC 
by operating 6.3-volt filaments at 5.95 volts. Later tests indi- 
cated that it may be better to use a value between 6 volts and 
6.3 volts. The filaments are left turned on for as long a period of 
time as possible, since several tube failures usually result from 
turning the equipment on. A large number of failures have 
resulted from open heaters, damaged cathodes, and near shorts 
in the tube framework. 3 It was found from examination of all 
the tubes which failed that the failure rate was considerably 
higher in tubes with folded filaments than in tubes with twisted 
filaments. Experience also indicates that low-voltage filaments 
last longer than high-voltage filaments of the same power rating. 
For this reason manufacturers provide series filaments in twin 
triodes of the industrial variety which are rated for a life of 
10,000 hours. 4 

High plate-load resistances are used in the ENIAC in order 
-to minimize the effect of changes in tube plate resistance. The 
average tube has lost emission noticeably after approximately 
9,000 hours of operation, but actual failure by this cause can be 
minimized by replacing tubes periodically. It has been suggested 
that provision be made in a computer to vary the voltage of the 



FACTORS AFFECTING CHOICE OF EQUIPMENT 433 

power supplies over a fairly wide range periodically. In this 
manner a large percentage of the incipient failures in tubes 
and other components could be made to occur. 

In the ENIAC the operating frequency is varied, but not the 
voltages. 

The practice of replacing a large number of tubes at once has 
the disadvantage that it may complicate subsequent trouble 
shooting inasmuch as most of the tube failures occur either within 
the first 250 hours of operation or after the 5,000th. The present 
ENIAC operating staff is not in sympathy with a policy of 
replacing a large number of tubes at once. 

Characteristics other than those affecting the life expectancy 
of electron tubes are equally important in considering the 
reliability of these components. For example, in applications 
where low-level voltages are to be amplified, as in the amplifiers 
for the output of magnetic recording pickups, the microphonic 
characteristics of tubes must be considered by the designer. 

Crystal diodes have been tested by the Eckert-Mauchly Com- 
puter Corporation under conditions typical of usage in electronic 
computers. The results of these tests have shown that the 
manufacturers' tests under other conditions are not necessarily 
valid for computer use. 

17-2-2. Capacitors. The capacitor is an electronic component 
which is subject to deterioration and eventual failure. 1 The 
length of its life depends upon the applied voltage, the tempera- 
ture, and the degree to which it is suited to its application. A 
statistical rule has been determined which shows that the mini- 
mum life to be expected of a group of capacitors before a large 
number fails is proportional to some power of the ratio of the 
applied voltage and the rated voltage. This exponent varies 
between 4 and 6 for conventional capacitors. The minimum 
life expectancy of capacitors is also a function of ambient tem- 
perature which should be carefully estimated before voltage 
ratings are established. 

17-2-3. Resistors. The life expectancy of a resistor with a 
given power rating depends upon the actual power dissipated 
and the ambient temperature. The small resistors used in the 
ENIAC are operated at approximately one-half rated power 
and at a temperature of 110°F, as compared to about 100°F, 
on which the rating is based. 2 Wire-wound resistors are operated 



434 PHYSICAL COMPONENTS AND METHODS 

at about one-fifth rated value in order to decrease their operating 
temperatures, and consequently the temperature to which nearby 
equipment is subjected. 

About 70,000 resistors are employed in the ENIAC, and not 
more than five were known to have caused a machine failure in a 
9,000-hour period of operation. 

17-2-4. Fuses and Indicators. An individual fuse was used 
for each power-supply voltage in the ENIAC. The fuses are 
made by Western Electric and have a small switch which closes 
when a fuse opens. These switches are used to turn off all high 
voltages when one fuse goes out. In the IBM Automatic 
Sequence Controlled Calculator, neon lights are used to indicate 
shorted capacitors, open resistors, and low emission in tubes. 

17-3. Possibilities for Checking 

The ultimate goal in computer design is complete intrinsic 
reliability. Checking, or the detection and location of errors, 
should therefore be considered a means and not an end in itself. 
Usable solutions to problems are obtained only when the equip- 
ment is operating properly, and no amount of checking equipment 
will procure a correct solution. Errors in computation may be 
machine errors or operator errors, and checks are required for 
each. The amount of checking equipment and the frequency 
with which checks are programmed into the machine depend 
primarily upon the reliability expected of the machine. When 
errors are detected in the results, it is necessary to locate the 
source of these errors and to eliminate their causes. The detec- 
tion and location of errors are discussed in this section with 
respect to both machine and human errors. 

17-3-1. Machine Errors. A machine error may be caused by 
either of two types of equipment failure: (1) an intermittent, or 
(2) a persistent failure. The detection of components which 
fail intermittently in conventional equipment may be very 
difficult. In relay computers, for example, approximately 85 per 
cent of the failures are intermittent. The majority of the failures 
in the ENIAC also have been of the intermittent type. 

Efficiency of operation of the ENIAC over any given time 
interval is closely related to the number of intermittent failures 
occurring during that interval. 



FACTORS AFFECTING CHOICE OF EQUIPMENT 435 

It has been estimated that the ENIAC is actually in useful 
operation on problems about 50 per cent of the time and that, 
allowing for a complete repetition of each problem, it therefore 
turns out about one-fourth as many useful answers as could an 
ideal, perfectly reliable, machine with the same basic mathe- 
matical characteristics, which required no maintenance what- 
ever. This is an excellent performance record. 

In the Bell Telephone Laboratories machines, continuous auto- 
matic checking is done at the level of the individual relay. This 
is called low-level checking. In the ENIAC, checks are made of 
solutions and partial solutions to problems, and this is termed 
high-level checking. When a failure occurs in a Bell Telephone 
Laboratories machine, all digital results are held in their respec- 
tive relays until the trouble is cleared, whereupon the computer 
continues to operate with the results obtained previously. 
Operation of the ENIAC is so fast, however, that it was not felt 
worth while to provide similar holding equipment, and results 
are recalculated after a failure. 

Several types of high-level checks have been worked out for 
computing equipments, 7 and they will be described briefly as 
follows: 

1. If a problem can be solved in two different ways in a com- 
puter, agreement between two answers so obtained con- 
stitutes a high-level check on the correctness of the solution. 

2. A check on the equipment can be made by getting answers 
to the problem in some way other than using the same 
equipment. For example, part of a solution could be 
worked out on a desk calculator, or perhaps solutions 
to restricted cases might be found and checked against 
the machine solution. 

3. In physical problems, the results are usually continuous, 
and the machine can be programmed to check high-order 
differences and cause an alarm to signal the operator if 
these vary too much. 

4. Special problems with known results may be designed to 
test certain sections of the machine. In this manner the 
machine may be stopped at various stages of the solution 
and the machine results compared with the known results. 



436 PHYSICAL COMPONENTS AND METHODS 

The machine operation at which the first error is made 
can be located in this manner. 
5. To check for intermittent errors, it may be sufficient to 
repeat the same problem two or more times. The results 
of the two runs may be checked in a comparator and the 
presence of any discrepancy detected. 

The ENIAC has indicator lights to represent the digits in each 
register. There is also provision for supplying the computer with 
one clock pulse at a time. Thus, an operator can run off the 
problem at low speed and observe the effect of each pulse on the 
quantities in the register, and the sources of error can be located 
quite accurately. This step-by-step operation is not feasible 
in a machine which uses dynamic storage mediums, since these 
generally depend upon a constant high-frequency pulse rate in 
order to maintain synchronism. 

The advantages of a machine which handles data in a serial 
fashion have been cited previously. Since there is less equip- 
ment, errors due to a failure are likely to show up soon after the 
failure occurs, and the number of steps required to locate the 
source of the trouble thereafter is usually small. On the other 
hand, if the serial machine has dynamic storage, it cannot be 
slowed down readily for the step-by-step check just described. 
In any case, a systematic program of periodic checks is desirable 
so that errors which are not detected by low-level checks will be 
found within a reasonable time after they occur. 

Perhaps the best example of a low-level check is that incor- 
porated in the Bell Telephone Laboratories machines, which 
employ a biquinary representation of numbers. 8,9 In this 
representation, one group of five relays indicates the digits from 
through 4, and a second set of two relays represents the digits 
and 5. Thus each decimal is represented by one relay in each 
set, which allows a simple check to be made on the operation of 
the relays. Each one has extra contacts which are connected 
in an alarm circuit so that they indicate any situation in which 
more or less than one relay in each pair of sets is actuated simul- 
taneously. Thus, the failure of one relay to operate because of 
dust between the contacts of another relay is detected. The 
operation of a relay which is not supposed to operate is also 
detected. These situations cause the equipment to stop operat- 



FACTORS AFFECTING CHOICE OF EQUIPMENT 437 

ing until the trouble is cleared. The biquinary check fails when 
two compensating failures occur simultaneously, but this possi- 
bility proves to be very remote. 

Another possibility for low-level checking is to provide two 
identical accumulators which operate on the same digits simul- 
taneously. Anticoincidence circuits note any difference in 
condition between the two accumulators and stop the computa- 
tion if such differences occur. This duplication might be fairly 
economical if the extra accumulator were used for shifting to the 
right or to the left in the accumulator, so that the shift is accom- 
plished by transferring the number from accumulator 1 to 
accumulator 2 and readjusting accumulator 1 to agree with 
accumulator 2. 

Duplicate systems and anticoincidence circuits may also be 
used at other critical units in order to detect incorrect operation. 

In the computer proposed by the Raytheon Company, 6 self- 
checking and diagnostic equipment is provided throughout. 
It is estimated that the proposed checking system demands the 
use of about 20 per cent additional equipment. The checking 
covers all storage and transfer of words, all arithmetic opera- 
tions, all selections made by the memory, and most other machine 
processes. 

For example, to check the storage and transfer of numbers, 
R. M. Bloch has devised a system employing a check number 
associated with each number. This check number is merely a 
weighted sum of the binary digits in the number. When a new 
number is constructed in the arithmetic unit, or is introduced 
into the machine, it is provided with a check number, which is 
then stored with it. When the original number is withdrawn 
from storage, a new weighted sum is computed and compared 
with the old one which was stored with it. Any discrepancy 
between the new and old weighted sums indicates an error made 
during storage or transfer. These check numbers can also be 
used to check arithmetic operations. 

17-3-2. Human Errors. Perhaps the ultimate computing 
machine will take data from measuring devices, automatically 
convert them to the desired form for the computer, perform the 
required computations, and deliver the written results in the 
optimum written form. Even though the operator is eliminated 
as an intermediate handler of the data, instructions must still 



438 PHYSICAL COMPONENTS AND METHODS 

be prepared for the computer so that it will perform the proper 
operations These instructions are first formed in terms of the 
mathematical steps desired and finally are written on a per- 
manent storage medium for automatic introduction into the 
computer. The intermediate process of expressing the mathe- 
matical steps in signals to the computing machine is known as 
coding. Coding may become a rather long and tedious process 
for some problems, and it may be advantageous to employ small 
computers to do the detail work in coding problems for the larger 
machines. In any case, it is possible to compare one coded 
medium with another, but the original assignment of instruc- 
tions must be checked either mentally or by experimenting on a 
sample problem whose solution is known. 

Two methods have been proposed for checking the work of the 
operator in writing the data and instructions on the input 
mediums. 5 As an example, consider the preparation of tapes 
with a keyboard. It is possible for two operators to prepare 
tapes at the same time. These two tapes are compared auto- 
matically, and only those marks which are identical on the two 
are admitted to the machine. Any difference in the two tapes 
stops the machine and signals the two operators, who must 
resolve the difference before entry of data into the machine can 
continue. 

In a slight variation of this idea, one tape is made up in 
advance, and an operator simulates the preparing of the tape a 
second time. As he operates the keyboard the second time, a 
typed record of his work is prepared for visual inspection. His 
typing is also automatically compared with the data on the 
original tape, and all identical data are inserted in the final tape. 
A discrepancy in the two stops the machine and indicates the 
error to the operator. Since he has the manuscript before him, 
the operator can compare it with what he has just typed and 
immediately resolve the difference. 

17-3-3. Location of Source of Errors. The sources of errors 
in computing equipment may be located with the aid of several 
of the devices which have been discussed in the sections on 
reliability and checking machine errors. For example, indicating 
fuses, neon lamps in resistor and capacitor circuits, and neon 
lamps to indicate the quantities in registers are all valuable aids 
to a technician who is thoroughly acquainted with the functioning 



FACTORS AFFECTING CHOICE OF EQUIPMENT 439 

of the computer. After troubles have been traced to one unit, 
it can be removed from the computer and subjected to diagnostic 
circuit tests if required. 

17-4. Size, Cost, and Speed 

Table 17-1 shows the number of digits which can be retained 
in various storage mediums for a unit cost and in a unit volume. 

Table 17-1 









Maximum convenient 










speed, digits per second 






Digits per 


Digits per 








Medium 


dollar 


cubic foot 






Remarks 
















Recording 


Reading 




Electromechanical 












relays* 


0.2-1 


300 


100 


100 




Electronic relays*. 


0.2-1 


300 


10 7 


10 7 




Punched cards .... 


2.5 X 10 5 


9 X 106 


2.5 X 103 


5 X 10 3 


10 3 binary digits per 
card, 0.4 cents per card 


Teletype punched 


3 X 10« 


3 X 10 7 


50 


50 


Mechanical sensing, 


tape 










0.05 cents per foot 


Teletype punched 


3 X 106 


3 X 10 7 


50 


7 X 10 3 


Optical sensing at 10 


tape 










feet per second, 0.05 
cents per foot 


Thirty-five-milli- 


6 X 10* 


3 X 10 7 


100 


10* 


Mechanically punch- 


meter film, 










ed, optically read at 


punched 










10 feet per second, 
2 cents per foot 


Wide punched tape. 


105 


2 X 10 7 


6 X 10 2 


6 X 10* 


Mechanically punch- 
ed, optically read at 
10 feet per second 


Thirty-five-milli- 


108 


10 9 


6 X 10 5 


6 X 10 5 


Optical sensing of 0.01- 


meter photo- 










by 0.02-inch spots, 


graphic film 










10 feet per second, 
5 cents per foot 


Magnetic wire 


2.5 X 10« 


10io 


10* 


10* 


0.004-inch diameter, 
0.010-inch bauds, 10 
feet per second, 0.05 
cents per foot 


Magnetic tape 


10 7 


3 X 10 9 


105 


105 


J<4-inch tape, five par- 
allel tracks, 0.020- 
inch digit spaces, 10 
feet per second, 0.1 
cent per foot 


Mercury delay line. 




2 X 10* 


lO" 


1 X 106 




RG-64U cable 




5 


106 


5 X 106 




Selectron* 




103 


10* 


5 X 10 3 




Deflection cathode- 






ray tube* 




102 


105 


2.5 X 105 





* Parallel storage; access time equal to reading time plus writing time. 



440 PHYSICAL COMPONENTS AND METHODS 

An approximation of practical reading and writing speeds is 
also given. The figures shown in this table are by no means 
precise, but may be of value in making order-of-magnitude 
estimates. 

REFERENCES 

Reliability of Components 

1. Brotherton, M., Capacitors (D. Van Nostrand Company, Inc., New 
York, 1946), pp. 45-55. 

2. Eckert, J., Jr., Reliability of Parts in Theory and Techniques for the 
Design of Digital Computers, Vol. II (Moore School of Electrical Engi- 
neering, University of Pennsylvania, Philadelphia, 1948), pp. 20-1- 
20-13. 

3. Michael, F. R., Tube Failures in ENIAC, Electronics, Vol. 20, No. 10, 
pp. 116-119 (October, 1947). 

4. Radio Corporation of America, Special Red Tubes (RCA Tube Depart- 
ment, Harrison, N.J., 1948). Pamphlet. 

Possibilities for Checking 

5. Alexander, S. N., Input and Output Devices for Electronic Digital Calcu- 
lating Machinery, Annals of the Computation Laboratory of Harvard 
University, Vol. XVI, pp. 248-253 (1948). 

6. Bloch, R. M., R. V. D. Campbell, and M. Ellis, The Logical Design 
of the Raytheon Computer, Mathematical Tables and Other Aids to Com- 
putation, Vol. Ill, No. 24, pp. 286-295 (October, 1948). 

7. Eckert, J. P., Jr., Reliability and Checking in Theory and Techniques 
for the Design of Digital Computers, Vol. IV (Moore School of Electrical 
Engineering, University of Pennsylvania, Philadelphia, 1948), pp. 
35-1-35-16. 

8. Williams, S. B., Reliability and Checking in Digital Computing Systems 
in Theory and Techniques for the Design of Digital Computers, Vol. IV 
(Moore School of Electrical Engineering, University of Pennsylvania, 
Philadelphia, 1948), pp. 34-1-34-16. 

9. Williams, S. B., Bell Telephone Laboratories Relay Computing System, 
Annals of the Computation Laboratory of Harvard University, Vol. XVI, 
pp. 41-53 (1948). 

Size, Cost, and Speed 

10. Sheppard, C. B., Transfer between External and Internal Memory, 
Annals of the Computation Laboratory of Harvard University, Vol. XVI, 
pp. 267-271 (1948). 



INDEX 



Abacus, 4 

Aberdeen, Md., Ballistic Research 
Laboratories, 191, 194-195, 201, 
214-215 
Abramowitz, M., 102 
Access time, defined, 304 
Accumulator, counter-type adder as, 
269 
denned, 293 
ENIAC, 52-54 
parallel ring, 293-294 
Accuracy, defined, 224 
Adams, method of, 127-128 
Adder, algebraic, 280-285 

binary, amplitude-type, three-in- 
put, 286 
two-input, 285-286 
coincidence-type, three-input, 
276-280 
two-input, 269-276 
coded decimal, 289-293 
decimal, 45-48 
diode matrix, 293 
half, 270-272, 288-289 
parallel, 49-54, 293-294, 297-299 
RCA computron tube, 293 
serial, 45-49, 272-280, 294-295 
ternary, 287-289 

types of, denned, 269-270, 285, 
293 
Addition, algebraic, 283 
by analog, electronic, 238 

mechanical, 225 
binary, 271-276 
on desk calculators, 136-137 
excess-3 coded decimal, 290 
on punched-card equipment, 153, 

158, 162 
ternary, 287-288 



Addition table, binary, 80 
octal, 80 
ternary, 287-288 
Aiken, Howard H., 123, 183, 200 
Aitken's method of interpolation, 

108-109 
Alcohol, sonic properties of, 346 
Algebraic equations, solution of, by 
analog methods, 240, 242-245 
numerical, 117-126 
Algebraic sign, representation of, 15, 

283-285 
Amplifier, Ampec, 425-426 
transistor, 422-423 
traveling-wave tube, 421 
Analog computer, defined, 4, 223 
Analog computers, 4, 223-245 
electrical, 238-245 
electronic, 237-240 
mechanical, 225-237, 242-243 
And circuit, 37 
Anti-Rossi circuit, 44-45 
Armstrong frequency discriminator, 

423 
Automatic sequence-controlled cal- 
culator, denned, 63 
Automatic Sequence Controlled Cal- 
culator, IBM, 182-185 
(See also Harvard computer, 
Mark I) 

B 

Ballistic Research Laboratories, Ab- 
erdeen, Md., 191, 194-195, 201, 
214-215 

Bargmann, V., 100 

Bateman, H., 100, 126 

Beard, R. E., 237 

Bell Telephone Laboratories (BTL), 
182, 187-191, 391-394, 401, 
421-422, 435-436 



441 



442 



INDEX 



BTL computers, Model I, 187 

Model II, 188 

Model III, 188 

Model IV, 188 

Model V, 182, 188-191 

Model VI, 182, 191 
BTL converters, data, 390-394 

speech, 401 
BTL transistor, 422-423 
BTL traveling-wave tube, 421 
Bennett, A. A., 100, 126 
Berkeley, E. C, 5 
Berry, C. E., 244 
Bessel's formula for interpolation, 

106-107 
BINAC, 182, 203, 214, 348 
Binary coded decimal system, 205, 
289-290 

conversion to decimal, 413-414 
Binary system, 80-82, 91-94 

conversion to decimal, 78-79, 90 

conversion to octal, 42-43, 83 
Biot, M. A., 100 

Biquinary system, 76, 189, 436-437 
Bloch, R. M., 65, 437 
Bratt, J. B., 237 
Brown, G. S., 403 
Brown, G. W., 244 
Brown, S. L., 245 
Brunetti, Cledo, 424 
Buckingham, R. A., 237 
Buffer, 44-45 
Bureau of Standards, National 

(NBS), 202, 214-215, 354 
Burks, A. W., 62, 65, 91, 96, 201 
Burroughs Adding Machine Co., 135 

desk calculators, 140 
Bush, V., 228, 230 



Calculating Punch, IBM, 153, 157- 

158, 162 
Caldwell, S. H., 228, 230 
California, University of, 214 

at Los Angeles (UCLA), 229, 
236-237 
Campbell, R. V. D., 65 



Capacitors, reliability of, 433 

as storage medium, 354 
Card-programmed Calculator, IBM 

Electronic, 165 
Card-reading equipment, 151-152 
Carmichael, R. D., 74 
Carry, 273-281, 290-294, 297-299 
double, 297 
end around, 298-299 
simultaneous, 297 
Checking, 205, 434-439 
in BTL machines, 435-437 
duplicate input method, 438 
in ENIAC, 435 

high- and low-level, defined, 435 
in SDC Raytheon computer, 437 
in UNIVAC, 205 
Christmas-tree circuit, 377 
Coded decimal system, excess-3, 205, 

289-293, 413-414 
Coding, defined, 10, 438 
Coincidence circuit, defined, 37 

high-speed, 379 
Collator, IBM, 153-155, 161-162 
Command forms, 10, 59-72 
Complement, 87-89 
Component, computing, denned, 3 
Computing services, IBM, 164 
Computron adder tube, 293 
Comrie, L. J., 135 
Consolidated Engineering Company, 

244 
Control elements, 10 
Conversion, data, 33, 395-416 
(See also Converter, data) 
radix, binary-to-octal, 42-43, 76, 
83-84, 86 
binary-to-decimal, 75, 78-79, 
412-413 
fractional numbers, 90 
decimal-to-binary, 75, 411 
octal-to-binary, 43, 76, 83-84, 86 
table, 75, 86 
theorem, 78 
proof of, 96-98 
Converter, data, analog-to-digital, 
386-393 
curve-following, 402-404 



INDEX 



443 



Converter, data, microfilm, 389-390 
photographic-film type, 388- 

390 
p.c.m., 390-393 
wind-tunnel, 151 
digital-to-analog, 393-396 
in MIT differential analyzer, 

395 
p.c.m., 393-394 
rotation type, 394-395 
Selectron, 395-396 
parallel-to-serial, 268 
radix, 42-43, 410-414 
binary-to-octal, 42-43 
decimal-to-coded-decimal, 413- 

414 
octal-to-binary, 43 
serial-to-parallel, 268 
speech, 401 
Counter, electromechanical, 12-13, 
388 
electronic, binary, 13-19 
pentode flip-flop, 25-26 
with single input, 15 
triode flip-flop, 14 
with two inputs, 16 
decade, 26 

modified-binary, 27-28 
pentode flip-flop, 26 
scaler, 27, 29 
ENIAC, 23-25 
ring, as adder, 293-294 
pentode flip-flop, 26 
thyratron, 21-22 
triode, 20-21, 23-25 
modulo r {see Counter, electronic, 

ring) 
rotation, photographic film, 388- 
390 
stripped gear, 387-388 
Veeder-Root, 388 
Counting, on punched-card equip- 
ment, 153, 158, 160-161 
Counting systems, 74-77 
Counting theorems, 77-79 
Crystals, reliability of, 432-433 
Curtis, R. W., 424 
Curve-followers, 229, 402-404 



Darmstadt, Technische Hochschule, 

403 
Decade counter, pentode pair, 26 

triode pair, 23-25 
Decade ring, pentode, 26 

triode, 23-25 
Decade scaler, 27 
Decimal adder, 45-48 
Decimal counter, 23-29 
Decimal point, floating, 91, 185, 189 

in radix conversion, 89-90 
Decimal system, 74-79, 90, 411-413 
Delay, defined, 303 
Delay line {see Storage, delay line) 
Descartes' law of signs, 124 
Desk calculators, 135-145 

Burroughs, 140 

Friden, 140-141 

manufacturers of, 135 

Marchant, 141 

Monroe, 142 

operation of, 136-138 

Remington Rand, 142 
Difference tables, 102, 105 
Differential analyzer, 230-240 

electronic, 237-239 

G.E., 236-237 

MIT, 230-236 

UCLA, 236-239 
Differential equations, 126-130, 240 

solution of, numerical, 126-130 
with network analyzer, 240 
Differential gears, 225 
Differentiation, by analog, elec- 
tronic, 238-239 " 
mechanical, 227-229 

numerical, 109-111 
Digital computer, defined, 3 

table of current U.S. projects, 
214-215 
Division, by analog, mechanical, 228 

approximate formula for, 60-61 

binary, 82-83, 91-94 

on desk calculators, 137 

octal, 83 



444 



INDEX 



Division, on punched-card equip- 
ment, 153, 158, 162 
Durfee, B. M., 183 
Dvorak, A., 398 



E 



Echoes, in sonic delay lines, 344 
Eckert-Mauchly Computer Corp., 

203, 214, 433 
EDVAC, 200, 215 

Elementary Potential Digital Com- 
puting Component (EPDCC), 
defined, 12 
Ellis, M., 65 
End-around carry, 89 
Engineering Research Associates, 
Inc. (ERA), 91, 208-213, 336- 
338, 407 
ENIAC, 182, 194-198, 214 
accumulator, 52-54 
addition in, 52-54 
decade counter, 23-25 
operating history of, 432-438 
plus-minus indicator, 15 
pulse transfer in, 375-377 
subtraction in, 51 
Equations, algebraic, 117-126, 240, 
242-245 
differential, 126-130, 240 
simultaneous, 125-126, 242-245 
transcendental, 117-120 
Euler-Maclaurin formula, 112 
Excess-3 code, 205, 289-293, 413- 

414 
Exponential function, generation of, 
230 



Failure of components, indications 
of, 434, 438-439 
intermittent, 434 
persistent, 434 
False position, method of, 119 
Federal Telephone and Radio Corp., 
352, 354, 391 



Feedback, in electronic differential 

analyzer, 239 
Feller, W., 108 
Fixed-cycle operation, 269 
Flip-flop, 13-17, 25-26, 28, 312 

cathode-ray tube, 28 

pentode, 25-26 

pulse-sensitive, 17 

operating speed of, 312 

triode, 14 

tube characteristic inversion, 28 
Floating decimal point, 91, 185, 189 
Ford, L. R., 100 
Franklin Corporation, 426 
Frequency-discrimination equip- 
ment, 423-424 
Friden Calculating Machine Co., 
Inc., 135, 140-141 

desk calculator, 140-141 



Gate, 37-42, 377-379 

common cathode, 39 

defined, 37 

diode, 41-42, 379 

dual grid, 37-38 

inverse, 44-45 

multiple, 377-379 

parallel, 38 

resistance grid, 40 

Rossi, 38 
Gauss, formula of, for numerical 

integration, 113, 115 
Gears, differential, 225 
General Electric Co. (G.E.), 236, 

240, 354, 406 
G.E. differential analyzer, 236-237 
General-purpose computer, defined, 
252 

factors affecting choice of, 260-261 
Globe Wireless Co., 400-401 
Globe Union Co., Centralab Div., 

425 
Goldberg, E. A., 244 
Goldstine, H. H., 62, 65, 91, 96, 201 
Graeffe, root-squaring method of, 
123-125 



INDEX 



445 



Haeff, Andrew V., 365 
Haeff storage tube, 365-366 
Half-adder, 270-272, 288-289 
Haloid Corporation, 322 
Hamilton, F. E., 183 
Harmonic analysis, 245 
Hartree, D. R., 5, 230, 237 
Harvard University Computation 

Laboratory, 13, 200, 214, 406 
Harvard computer, Mark I, 13, 35- 
36, 182-185, 214, 434 

description of, 182-185 

indication of failures in, 434 

relays used in, 35-36 
Mark II, 36, 182, 185-187, 214, 
311-313 

description of, 185-187 

relays used in, 36, 311-313 
Mark III, 182, 200, 214 

description of, 200 
Hazen, H. L., 403 
Hermite, interpolation formula of, 

109 



Illinois, University of, 215 
Impedance, acoustic, 343 
Input, defined, 6 

Input equipment, 6-7, 150-152, 
396-404 
punched card, 150-152 

(See also descriptions of vari- 
ous computers) 
Input table, for differential analyzer, 

229 
Institute for Advanced Study, 

Princeton, N.J., 201-203, 215 
Integration, by analog, electronic, 
237-239 
mechanical, 227-230 
numerical, 111-116 
Integrator, inverse, 228 

wheel-and-disk, 227-230, 232, 236 
Interfiling Reproducing Punch, Rem- 
ington Rand, 158-159 



International Business Machines 

Corp. (IBM), 13, 147, 151, 163, 

312 

IBM accounting machines, 153-158, 

160-165 

applications of, to scientific work, 

164 
Collator, 153-155, 161-162 
Multiplier, 153-156 
Punch, Calculating, 153, 157 
Electronic, 153, 157-158 
Reproducing, 162-163 
Punched Hole Verifier, 151 
Sorter, 153-154, 160-161 
Tabulator, 153, 155-156, 160, 162 
IBM punched card, 148-150 
IBM sequence-controlled calcula- 
tors, Automatic Sequence Con- 
trolled Calculator, 182-185 
(See also Harvard computer, 
Mark I) 
Card-programmed Electronic Cal- 
culator, 165 
Pluggable Sequence Relay Calcu- 
lator, 182, 191-193 
Selective Sequence Electronic Cal- 
culator, 182, 198-200 
Interpolation, 100-109 
Aitken's method, 108-109 
Bessel's formula, 106-107 
defined, 100 
inverse, 108 

Lagrange's formula, 107-108 
Newton's formulas, 103-104 
by proportional differences, 108- 

109 
Stirling's formula, 104-106 
trigonometric, 109 
Inverse gate, 44-45 
Inverse, iteration methods for find- 
ing, 122-123 
Isolating circuits, 44-45 



Jaeger, J. J., 403 

Jump, 67 

Justo writer Company, 401 



446 



INDEX 



K 



Kallmann, H. E., 349-352 

Kelvin wheel-and-disk integrator, 

227-230 
Keyboards, 397-399 

Burroughs, 397-398 

Dvorak, 398 

stenotype, 398-399 

Sundstrand, 397 

teletypewriter, 398 
Kilburn, T., 366 
Klystron, 421-422 



Laderman, J., 102 

Lagrange's formula for interpolation, 

107-108 
Lake, C. D., 183 
Lennard-Jones, J. E., 237 
Linear mapping, 94-95 
Link Paper Co., 317 
Logarithm, generation of, 230 
Logic, computing-machine, defined, 
62 
rules of, applied to arithmetic 
elements, 270-285 
Logical and circuit, 37 
Logical or circuit, 44-45 



M 



Magic-T, 423 

Magnetic drum, as cyclic storage 
system, 304-306 
ERA design standards for, 336- 

339 
recording {see Recording, mag- 
netic) 
speed changers, 415 
Magnesium, sonic properties of, 344, 

346 
Mallock's circuit, 243-244 
Manchester, University of, 354 
Marchant Calculating Machine Co., 
135 
desk calculator, 141 



Margenau, H., 100, 125 
Massachusetts Institute of Tech- 
nology (MIT), 214, 229-230 
MIT digital computer (Whirlwind 

I), 214 
MIT differential analyzers, 230-236 
MIT storage tube, 354, 364-365 
Massey, H. S., 237 
Matrix calculations, 126 
Matrix, diode, 40-43, 413-414 

resistor, 40-41 
Mayan system of arithmetic, 76-77 
Memory (see Storage) 
Mercury, sonic properties of, 306, 
342-348 
table of, 346 
storage systems, 306-308, 341-348 
(See also Storage, delay line) 
Millen, James, Co., 354 
Milne, method of, 129-130 
Modulation, pulse code (p. cm.), 

390-394 
Modulo, defined, 15 
Monoscope, 408 

Monroe Calculating Machine Co., 
Inc., 135 
desk calculator, 142 
Montgomery, D., 100 
Moore, B. L., 200, 339 
Moore School of Electrical Engineer- 
ing, University of Pennsylvania, 
194, 200, 214-215 
Morley, S. G., 76 
Morton, G. A., 369 
Multiharmonigraph, 226, 238, 245 
Multiplication, by analog, elec- 
tronic, 238 
mechanical, 226-228 
binary, 80-81 

table, 81 
on desk calculators, 137 
octal, 81-82 

table, 81 
on punched-card equipment, 153, 
158, 162 
Multiplier, IBM, 153, 156 
parallel, 299-301 
Remington Rand, 158, 160 



INDEX 



447 



Multiplier, serial, 295-297 

slide, 226 
Murray, F. J., 237 
Murphy, G. M., 100, 125 

N 

NACA, 214 

National Bureau of Standards 

(NBS), 202, 214-215, 354, 407 
Naval Proving Ground, Dahlgren, 

Va., 91, 185, 191, 214 
Naval Research Laboratory (NRL), 

computers, 188 
NRL storage tube, 354, 365-366 
Negative number, representation of, 

87-89 
Network analyzers, 240-242 
Newton-Cotes formula, 112 
Newton's formula, applied to numer- 
ical differentiation, 110 
for backward interpolation, 104 
for forward interp-^ation, 103-104 
Newton-Raphson method, 117-119, 

125 
Northrop Aircraft, Inc., 203, 341 
Number, denned, 302 
Number representation systems, bi- 
nary, 75-76 
binary-coded dscimal, 205 
biquinary, BTL, 76, 189 
British monetary, 77 
decimal, 74 
octal, 76 
table of, 86 

tetraquinary, Mayan, 76 
vigesimal, Mayan, 76-77 

(See also descriptions of various 
computers) 
Number theorems, 78-79, 89 

proofs of, 96-99 
Numerical analysis, 100-130 
Numeroscope, 406-407 



Octal system, 76, 80-83 
addition in, 80 
division in, 83 
multiplication in, 81-82 



Octonary (see Octal system) 
Off-line operation, 7, 260-261 
O'Neal, R. D., 319, 339 
On-line operation, 7, 260-261 
Optical recorder, 321 
Or circuit, 44-45 

Output equipment, 10-11, 229, 404- 
410 
differential analyzer, 229 
printing, counter-wheel, 410 
electric typewriter, 229, 405 
gang, 405-406 
photographic, 406-407 
Veeder-Root, 410 
punching, 229 
recording, 408-410 

(See also descriptions of various 
computers) 



Parallel-to-serial conversion, 267- 

268 
Pennsylvania, University of, 194, 

200 
Periodic functions, interpolation of, 

109 
Photographic film, as storage me- 
dium, 319-322 
Point, radical, 89-91 
Porter, A., 237 
Potted circuits, 426-427 
Pound, R. V., 423-424 
Precision, defined, 224 
Printed circuits, 424-426 
Printer, gang, 406 
IBM, 155-156, 406 
numeroscope, 406-407 
photographic, 406-408 
Remington Rand, 158-160, 406 
Program, example of, for automatic 
sequence-controlled calcula- 
tor, 69-71 
for desk calculator operator, 
58-62 
relationship of, to computer de- 
sign, 252-261 
Pulse code modulation, 390-393 
Pulse transfer, 374-379 



448 



INDEX 



Punched card, Hollerith, 146 
IBM, 148-150 
Powers, 150 
Remington Rand, 150 
Punched-card computing systems, 
146-181 
IBM, 151-158, 406 
manufacturers of, 147, 163 
Remington Rand, 152, 158-160, 
406 
(See also IBM; Remington 
Rand.) 
Punched Hole Verifier, Electric, 
IBM, 151 

Q 

Quartz, sonic properties of, 346 
R 

Rack, A. J., 394 
Radar, 305 
Radical point, 89-91 
RCA, electrical analog computer, 
244 
frequency-stabilization system, 

424 
miniature tubes, 422 
storage tube, 354 
Radix, choice of, 84-87, 420-421 
conversion, binary-to-decimal, 75, 
78-79, 412-413 

fractional numbers, 90 
binary-to-octal, 42-43, 76, 83- 

84, 86 
decimal-to-binary, 75, 79, 411 
octal-to-binary, 43, 76, 83-84, 86 
table, 75, 86 
theorem, 78 
proof of, 96-98 
defined, 80 
Rajchman, J. A., 363, 370 
Raydist, 253-260 

Raytheon Mfg. Co., adder, 297-298 
computer, general description, 
206-208 
number representation in, 267 
miniature tubes, 422 
storage tube, 354 



REAC, 239 

Recording, magnetic, 322-339 

digital, compared with sound, 
325-326 
design considerations, 334- 

339 
ERA, 336-338 
number representation in, 

328-332 
number transfer in, 332-333 
systems, 333-334, 336-339 
transfer of digits in, 332-333 
erasure of, 326-328 
heads, 324-326 
longitudinal, 325 
mediums, 323-324 
perpendicular, 325 
process, 323-326 
Reeves, A. H., 391 
Reeves Instrument Corporation, 239 
Register, defined, 295 
BTL, 188-190 
shifting, 297-299 
Relay, electromechanical, 35-37, 
235-236, 304, 311-314 
wire brush, 312 
electronic, 13-17, 25-26 
(See also Flip-flop) 
Relay computers, 182-193 
Remington Rand, Inc., 135, 142, 

147, 163 
Remington Rand desk calculator, 

142 
Remington Rand punched card, 150 
Remington Rand punched-card ma- 
chines, 152, 158-160, 163 
Punch, Alphabetical, 152 

Interfiling Reproducing, 158- 

159 
Printing Multiplier, 158, 160 
Reproducing, 152 
Summary, 152 
Synchro-Matic, 152 
Sorter, 158-159 
Tabulator, 158-160 
Resistors, reliability of, 433-434 
Ritt, J. F., 89 
Robinson, G., 100, 125, 126 



INDEX 



449 



Roots of equations, by method of 
Graeffe, 123-125 

by numerical methods, 117-126 
Rossi circuit, 37-38, 379 

high-speed, 379 
Round-off error, 95-96 
Runge-Kutta, method of, 128-129 



Saturable-core reactor, 310 

Scarborough, J. B., 100, 125, 126 

Schrodinger equation, 241 

Screw differential, 225 

Secondary emission, 355-361 

Secondary-emission ratio, 356 
effective, 355-356 

Seeley-Foster frequency discrimina- 
tor, 423 

Selective Sequence Electronic Calcu- 
lator, IBM, 182, 198-200 

Selectron, 308, 370-374, 395-396, 
408 

Sequence-controlled calculator, de- 
fined, 63 

Serial-to-parallel conversion, 267- 
268, 303 

Shannon, C. E., 230, 394 

Shifting register, 297-299 

Sign, algebraic, representation of, 
15, 283-285 

Simpson's one-third rule, 112, 114 

Simpson's three-eighths rule, 112 

Simultaneous equations, 125-126 

Slide multiplier, 226 

Slide rule, 4 

Sonic delay lines {see Storage, delay 
line, sonic) 

Sonotone, 422 

Sorter, IBM, 153-154, 160-161 
Remington Rand, 158-159 

Sound, transmission properties of, 
306, 341-346 
table of, 346 

Special Devices Center (SDC) com- 
puter, 206-208 

Special-purpose computer, defined, 
252 



Special-purpose computer, factors 

affecting choice of, 260-261 
Speech converters, 401 
Speed-changing devices, 414-416 
Spinden, H. J., 76 

Square, generation of, by electronic 
circuit, 239-240 
by integrator, 229-230 
Square root, extraction of, by elec- 
tronic circuit, 239-240 
by numerical methods, 58, 120- 
122 
Stanford University, 350 
Stibitz, George, 76 
Storage, access time, 304 
cyclic, denned, 304-308 

mediums applicable to, 305-308 
cycle, defined, 304 
defined, 9, 302-303 
delay line, 341-354 

electromagnetic, 349-354 

artificial transmission line, 

353-354 
waveguide, 350, 354 
wire transmission line, 351- 
352, 354 
sonic, 304-308, 341-348 
aqueous solution, 345 
ethylene glycol, 342 
magnesium, 342, 344 
mediums, 341-342, 346 
mercury, 306-308, 341-348 
quartz, 344 

structure of, 342, 347-348 
transducers for, 342-346 
dynamic, defined, 305 
electrostatic, 354-374 
Haeff tube, 365-366 
Krawinkel tube, 369 
Manchester tube, 366-369 
MIT tube, 354, 364-365 
NRL tube, 365-366 
Selectron, 370-374 
Williams tube, 366-369 
Zworykin and Morton tube, 369 
erasable, defined, 305 
magnetic, 306, 322-339 
erasure techniques, 326 



450 



INDEX 



Storage, magnetic, digital signal 
types, 328-332 
system design considerations, 

334-338 
transfer schemes, 332-333 
(See also Recording, magnetic) 
mercury, 306-308 

(See also Storage, delay line, 
mercury) 
noncyclic, defined, 304 

mediums, applicable to, 308-310 
nonvolatile, defined, 305 
paper and pencil, 63 
permanent, defined, 305 
phosphor, 339-341 
photographic, 318-322 
punched tape, 314-319 
relay, 311-314 
static, defined, 304 
ternary, 312, 420 
volatile, defined, 305 

(See also descriptions of various 
computers) 
Subtraction, by analog, electronic, 
238 
mechanical, 226 
binary, 281-284 
decimal, 47-51 

on desk calculators, 51, 136-137 
on punched-card equipment, 153, 
158, 162 
Subtracter, binary, 281-284 
Sundstrand Keyboard, 397 
Switching, in noncyclic storage sys- 
tems, 308-310 
Sylvania Electric Products, Inc., 

406, 422 
Synchro-Matic Punch, Remington 
Rand, 152 



Table consulting, on punched-card 

equipment, 153-155, 158, 161 
Tables, mathematical, on IBM cards, 

164 
Tabulator, Alphabetic, IBM, 159 
IBM, 153, 155-156, 160, 162 
Remington Rand, 158-160 



Tape-reading equipment, 151, 317- 

318 
Teledeltos paper, 410 
Teleregister Corporation, 399 
Teletype, operation of, 399-400 

tape, as storage medium, 317-318 
Ternary system, 84-87, 287-288, 
312, 420 

advantages of, 84-87, 420 
Tetraquinary system, 76 
Theorems, number, 78-79, 89 

proofs of, 96-99 
Transcendental equations, 117-120 
Transducers, 342-348 

ADP, 344, 346 

quartz, 342-348 
Transfer of control, conditional, 67 

unconditional, 68 
Transient analyzer, 241-242 
Transistor, 422-423 
Trigonometric interpolation formula, 

109 
Tube, traveling-wave, 421 
Tubes, vacuum, reliability of, 432- 

433 
Tyler, A. W., 319, 339 
Typewriter, Electromatic, 400 



U 



Underwood-Elliott Fisher Co., 102 
UNIPRINTER, 205 
UNITYPER, 205 
UNIVAC, 203-206 
University of California in Los 
Angeles, 229, 236-237 



Vigesimal system, 76 
Von Karman, T., 100 
Von Neumann, John, 62, 65, 91, 96, 
100, 201 



W 



Watson Scientific Computing Labo- 
ratory, 191-193 



INDEX 



451 



Watson, T. J., 183 
Weddle's rule, 112, 115 
Weierstrass, theorems of, 101 
Wheeler, L. L., 245 
Whirlwind I computer, 214 

storage tube for, 354, 364-365 
Whittaker, E. T., 100, 125, 126 
Wilbur's mechanism, 242-243 
Wiener, Norbert S., 5 
Wilkes, M. V., 237 
Williams, F. C, 366 



Williams tube, 202, 366-369 
Word, 302 
Wylie, J., 237 



X 



Xerography, 322 



Zimmerman, K. H., 352 
Zworykin, V. K., 369 



a 



^Of